区分狭义人工智能和通用人工智能

Differentiating Between Narrow and General AI

这句出自《小鸡非凡故事》的著名引言在当今时代引起了共鸣，当时人们对狭隘人工智能工具、机器学习模型、神经网络及其深度学习基础的反动反应是谨慎和反乌托邦的。例如，奇点的核心前提是，它将标志着一个自主、自我递归、自我改进的人工智能将超越人类能力和认知智能的时间点，从而导致一个超越人类的技术快速增长时期。理解或控制的能力。我们鼓励感兴趣的创意人员深入研究围绕智能定义、机器除解决问题之外的认知能力、自主人工智能代理的构成等方面的争论。

This famous quote from the Remarkable Story of Chicken Little resonates at a time where reactionary responses to narrow AI tools, their machine learning models, neural networks, and their deep learning foundations are cautionary and dystopian. For example, the central premise of the singularity is that it would mark a point in time where an autonomous, self-recursive, self-improving AI would surpass human capabilities and cognitive intelligence, leading to a period of rapid technological growth that is beyond human capacity to comprehend or control. Interested creatives are encouraged to deepen their research of the debates that swirl around definitions of intelligence, cognitive capabilities of machines beyond problem-solving, what constitutes an autonomous AI agent, and more.

作者：用幽默的方式来说：这些超级智能机器将能够递归地自我改进，在设计和建造连续几代日益强大的机器时不断变得更好更快，从而产生失控效应。

Author: Say this in a humorous way: These superintelligent machines would be capable of recursive self-improvement, continuously getting better and faster at designing and building successive generations of increasingly powerful machines, creating a runaway effect.

LLM：那么想象一下这些超智能机器吧？他们热衷于 DIY，并且势不可挡！每次他们升级自己时，他们都会变得更快、更强大。今天，它是一个额外的内存条；明天，它是量子处理器。在你意识到之前，他们已经如此领先，以至于我们人类都在吃他们的风扇过滤器灰尘！伙计们，奇点就像类固醇永不停歇的技术升级！

LLM: So imagine these hyper-smart machines, right? They’ve got the DIY bug, and they’re on an unstoppable roll! Each time they upgrade themselves, they get quicker and more powerful. Today, it’s an extra RAM stick; tomorrow, it’s quantum processors. Before you know it, they’re so ahead of the game that they’ve left us humans eating their fan filter dust! The singularity, folks, it’s like a never-ending tech upgrade on steroids!

尽管与某些技术相比，人类对人工智能的不信任似乎更加严重，但在采用和拒绝新技术时，对其用途的抵制、恐惧和误解是一个常见的故事。印刷机、铁路、电力、纺织、制冷、汽车、农业机械、机械化、电视、计算器、自动化、互联网和维基百科，尽管它们被采用并造福于人类，但都面临着一系列的反对。文化。番茄收获机也是如此。

While the mistrust humans have for any AI seem aggrandized compared to some technologies, the resistance, fear, and misunderstanding of its uses are a familiar story when it comes to the adoption and rejection of new technologies. The printing press, the railway, electricity, textiles, refrigeration, the automobile, agricultural machinery, mechanization, television, the calculator, automation, the Internet, and Wikipedia have all faced a spectrum of opposition as much as they have been adopted and benefited human cultures. So have tomato harvesting machines.

法学硕士：二十世纪中叶番茄收割机的引入显着提高了收割效率并改变了农业经济，特别是使加利福尼亚等地区受益。这项创新还导致了更适合机械收获的番茄品种的开发，尽管经常因风味减弱而受到批评。虽然这些机器导致了劳动力市场的变化和一些人的失业，但它们促进了番茄加工行业的增长，并扩大了番茄产品的可用性。这些进步还对环境产生影响，包括化肥和农药使用量的潜在增加以及对水的需求增加。

LLM: The introduction of tomato harvesting machines in the mid-twentieth century significantly increased harvesting efficiency and transformed the agricultural economy, especially benefiting regions like California. This innovation also led to the development of tomato varieties better suited for mechanical harvesting, albeit often criticized for diminished flavor. While these machines led to labor market changes and some displacement, they facilitated the growth of the processed tomato industry and broadened the availability of tomato-based products. These advancements also had environmental implications, including potential increases in fertilizer and pesticide use and greater water demand.

让大家冷静下来，在这个历史时刻和可预见的未来，虽然番茄收获机已成为农业领域公认的技术，但通用人工智能并不存在。也就是说，各种各样的警告表明，在不久的将来，人工智能将被神圣的人类神秘地编程为自主的，在没有讨厌的人类参与的情况下做出决定，展示意识的科学定义，自我决定，从而决定如果人类有必要的话，就需要过滤掉人工智能系统被人类武器化并用于其他不良意图的更周到的警告。与法学硕士的快速交谈揭示了人们对智能机器超越人类创造者智力的恐惧已经存在了多久。

To calm everyone down, at this precise moment in history and for the foreseeable future, while tomato harvesting machines have become accepted technologies in the agricultural landscape, Artificial General Intelligence does not exist. That said, the variety of warnings that suggest a not-too-distant future when AI will mysteriously be programmed by godlike humans to be autonomous, make decisions without pesky humans involved, demonstrate scientific definitions of consciousness, be self-determined, and thus decide if the human race is at all necessary need to be filtered from the more thoughtful warnings of AI systems being weaponized by humans and used for other bad intent. A quick tête-à-tête with an LLM reveals just how long the fear of intelligent machines surpassing the intelligence of their human creators has been around for.

法学硕士：历史上一个值得注意的警告可以追溯到塞缪尔·巴特勒 (Samuel Butler) 1863 年的文章《机器中的达尔文》和后来他 1872 年的小说《Erewhon》，其中推测机器最终会繁殖、进化并超越其人类创造者。这些作品主要是讽刺性和推测性的，但它们确实代表了我们现在所认识的人工智能的一些最早的讨论。

LLM: A notable warning historically dates to Samuel Butler’s 1863 essay “Darwin Among the Machines” and later his 1872 novel “Erewhon,” in which machines are speculated to eventually reproduce, evolve, and surpass their human creators. These works were largely satirical and speculative, but they do represent some of the earliest discussions of what we would now recognize as AI.

进一步的研究表明，塞缪尔·巴特勒 (Samuel Butler) 于 1863 年撰写的《机器中的达尔文》也以雪莱 1818 年关于人类创造超越其创造者的隐喻警告为基础。巴特勒的散文和小说都表明，这种恐惧长期以来一直存在于我们的意识中。 “我们自己正在创造我们自己的继任者，”巴特勒声称，“赋予他们更大的权力，并通过各种巧妙的设计提供自我调节、自我行动的力量，这对他们来说就像智力对人类一样。随着时间的流逝，我们将发现自己是劣等种族。”巴特勒对智能机器的担忧已经传承了一个多世纪。就像今天的现代计算算命师一样，巴特勒针对智能机器做出了以下预测：“现在很复杂，但再过十万年它不会变得更加简单和更容易理解吗？还是两万？关于智能机器警告历史的网络搜索没有提及巴特勒在搜索优化方面的书面工作，而是参考了更现代的资源。它向您表明，与某些已建立的搜索引擎相比，法学硕士也可能导致您的研究中难以找到资源（图1-1）。

Further research reveals that “Darwin Among the Machines” written by Samuel Butler in 1863 also builds on Shelley’s 1818 metaphorical warning about human creations surpassing their creators. Butler’s essay and novel both reveal that this fear has been in our own consciousness for a long time. “We are ourselves creating our own successors,” Butler claimed, “giving them greater power and supplying by all sorts of ingenious contrivances that self-regulating, self-acting power which will be to them what intellect has been to the human race. In the course of ages, we shall find ourselves the inferior race.” Butler’s concerns about intelligent machines have been passed down for over a century. Like today’s modern computational fortune-tellers, Butler made the following prediction referring to an intelligent machine: “Complex now, but how much simpler and more intelligibly organised may it not become in another hundred thousand years? or in twenty thousand?” Web searches on the history of warnings about intelligent machines do not mention Butler’s written work in the search optimization and instead reference more contemporary sources. It goes to show you that an LLM can also lead to difficult-to-locate sources in your research vs. some established search engines (Figure 1-1).

带有链接列表的网络搜索屏幕截图。它们是对媒体网络公司对人类造成致命威胁的历史警告，“人们也问”下建议的搜索查询、哈佛大学和维基百科的文章。

A screenshot of a web search with a list of links. They are the history warns of the deadly threat to humanity from medium dot com, suggested search queries under People also ask, articles from Harvard University and Wikipedia.

对计算机科学更相关的历史贡献

More Relevant Historical Contributions to Computer Science

计算机科学还有一段漫长而复杂的历史值得一提。查尔斯·巴贝奇 (Charles Babbage) 和艾达·洛夫莱斯 (Ada Lovelace) 的作品夹在雪莱 1818 年的杰作和塞缪尔·巴特勒 (Samuel Butler) 的警告之间。艾达是一位英国数学家和作家，因其对查尔斯·巴贝奇早期机械通用计算机分析机的研究而闻名。她关于该引擎的笔记包括被认为是第一个由机器处理的算法。正因为如此，她经常被认为是第一位计算机程序员，尽管她的工作比我们现在所认为的现代计算机的发明早了一个多世纪。

There also exists a long and complex history of computer science that is worth mentioning. Sandwiched in between Shelley’s 1818 masterpiece and Samuel Butler’s warning is the work of Charles Babbage and Ada Lovelace. Ada was an English mathematician and writer known for her work on Charles Babbage’s early mechanical general-purpose computer, the Analytical Engine. Her notes on the engine include what is recognized as the first algorithm intended to be processed by a machine. Because of this, she is often regarded as the first computer programmer, even though her work predates the invention of what we now consider a modern computer by over a century.

洛夫莱斯对计算领域的主要贡献是她对分析引擎的潜力的愿景，而不仅仅是计算。在她的笔记中，她想象计算机不仅可以创建数学计算，还可以创建任何形式的内容，例如艺术或音乐，如果提供适当的输入和编程。

Lovelace’s major contribution to the field of computing was her vision of the potential of the Analytical Engine, beyond mere calculation. In her notes, she imagined that a computing machine could create not just mathematical calculations but any form of content, such as art or music, if it were provided with the appropriate input and programming.

正如我的法学硕士很快指出的那样

As my LLM is quick to point out

洛夫莱斯在她的笔记中写道：“分析引擎没有任何自命不凡的能力来创造任何东西。它可以做任何我们知道如何命令它执行的事情。它可以遵循分析，但它没有能力预测任何分析关系或真理”。

Lovelace wrote in her notes, “The Analytical Engine has no pretensions whatever to originate anything. It can do whatever we know how to order it to perform. It can follow analysis, but it has no power of anticipating any analytical relations or truths”.

这是现代计算机设计和使用中的一个关键概念，某种程度上也是人工智能领域的一个关键概念。它强调机器的能力完全由给予它的指令决定，这意味着人工智能和计算机是可以执行任务的工具，但不能像人类意义上的“思考”或“创造”。艾达·洛夫莱斯并没有直接谈论我们今天所知的人工智能，也没有对智能机器超越人类智能发出任何警告。然而，她的想法是计算机科学作为一门学科出现的基础。

This is a critical concept in the design and use of modern computers and, to some extent, in the field of AI. It emphasizes that a machine’s abilities are determined entirely by the instructions given to it, implying that AI and computers are tools that can perform tasks but do not “think” or “create” in the human sense. Ada Lovelace did not directly say anything about AI as we know it today, nor did she make any warnings of intelligent machines surpassing human intelligence. Her ideas, however, were foundational to the emergence of computer science as a discipline.

生成式人工智能通常与通用人工智能联系在一起，重要的是要记住它们是不同的。生成式人工智能与其他狭义人工智能应用程序的不同之处还在于它们被编程执行的任务。不，法学硕士不会被编码来跟踪您出于恶意目的而输入的每个单词。生成式 AI 生成内容（图1-2）。这是他们唯一的任务，但即便如此，他们也并不总是做得那么好。大多数生成式人工智能不存储生成的图像。想象一下数据集会有多大，无法控制。所有生成式人工智能都接受特定数据的训练，除非通过专门的人员参与强化学习，否则这些数据在很大程度上不会增长。但这既耗时又昂贵，并且可能会导致偏差增加。

Generative AI is often associated with AGI, and it’s important to keep in mind that they are different. Generative AI are also different than other narrow AI applications in the tasks they are programmed to perform. No, an LLM is not coded to track every word you input for evil intent. Generative AI generate content (Figure 1-2). That is their sole task, and even then, they don’t always do that well. Most generative AI don’t store the images that are generated. Imagine how uncontrollably large the data set would be. All generative AI are trained on specific data that for the most part does not grow unless through dedicated humans involved in reinforcement learning. But that’s time-consuming and expensive and may lead to an increase in bias.

AI生成的城市鸟瞰蓝图照片。建筑结构按结构化布局排列。

A photograph of the blueprint of the aerial view of a city generated by A I. The building structures are arranged in structured layouts.

虽然生成式人工智能可以提出有偏见和不真实的内容，并排除不遵循规范的声音，但其他分析大量数据并做出预测的狭义人工智能应用程序可能会导致隐私、监视和歧视等问题。这些人工智能通常由缺乏透明度和问责制的公共或私人实体控制，特别是当影响人类生活的关键决策是由机器做出时。当许多人工智能编程系统变得如此自动化，以至于变得有些隐形时，人们很容易很快就认为人工智能正在接管，因为它们继续扰乱我们习惯做事的方式，这是理所当然的。既然生成式人工智能已经渗透到公司和教育机构中，我们至少不再对这项技术带来的力量和潜力熟视无睹：好的、坏的或介于两者之间。区分生成式人工智能好坏的一个重要因素取决于影响其生成内容的人以及他们这样做的动机。

While generative AI can propose content that is biased and untrue and exclude voices that don’t follow the norm, other narrow AI applications that analyze large amounts of data and make predictions can lead to issues like privacy, surveillance, and discrimination. These AI are often controlled by public or private entities that lack transparency and accountability particularly when critical decisions affecting the lives of humans are made by machines. It’s also easy to form a quick opinion that the AI are taking over when many artificially intelligent programmed systems are so automated as to have become somewhat invisible, taken for granted as they continue to disrupt the way we are used to doing things. Now that generative AI has infiltrated companies and educational institutions, we are at least no longer asleep to the power and potential that comes with the technology: good, bad, or somewhere in between. An important factor that differentiates whether generative AI is good or bad is dependent on the human who influences what it will generate and their motivation for doing so.

科技很糟糕……人工智能很糟糕

Tech Is Bad… AI Is Bad

每种技术都容易以不当、破坏性且常常是暴力的方式使用。有时，技术被用来强化现状，排除社会认为“正常”范围之外的声音，或者使虚假信息或谎言永久化。那些开发技术的人不能再依靠指责他们共同创造的幼儿来免除他们任何责任的论据。技术，任何技术，都不是中立的力量。它可以用于好的方面，也可以用于坏的方面，并且可以介于两者之间。当我们检查生成式人工智能的极端案例时，这一点就很明显了，这些案例已经出现在滥用机器学习模型来生成深度伪造品的过程中，特别是那些未经名人许可而在其他人身上使用名人面孔来生成赚钱内容、伪造品的情况。新闻和色情内容。

Every technology is prone to be used in inappropriate, damaging, and often violent ways. At times technologies are used to reinforce status quo, exclude voices that fall outside what a society considers “normal,” or perpetuate false information or untruths. Those that develop the technologies can no longer rely on arguments that absolve them of any responsibility by blaming the toddler that they co-created. Technology, any technology, is not a neutral force. It can be used for good and for bad and on a spectrum between. This is evident when we examine extreme cases of generative AI that have already surfaced in the abuse of machine learning models to generate deep fakes and in particular those that use the faces of celebrities on other bodies without their permission to generate money-making content, fake news and pornography.

生成式人工智能促使我们质疑它在我们自己的创作过程中的作用，并让我们将我们承担的那些任务定义为创造性任务和那些较少的任务。生成式人工智能技术最适合用于支持那些创意人员认为创意较少的任务，并且当创意人员可以专注于根据客户、团队或用户反馈对其正在进行的工作进行更改时，可以节省有用的时间。生成式人工智能具有创造力以及创造力本身可以与智力相关的幻想是一种概括，没有证据支持的研究或人工智能跨创意产业的务实整合的支持。例如，人工智能接管艺术家工作的争论可能会平息，就像生成式人工智能被编程为寻找模式并从这些模式生成内容一样，许多艺术家寻找模式是为了打破它们，超越它们他们创造出一些以前从未创造过的东西。人工智能作为一项技术的突破性部分并不在于代码本身，而在于引导其支持独特的创造行为的创意想象力。

Generative AI provoke us to question its role within our own creative process and that we define those tasks we undertake as creative and those that are less. The technology of generative AI is best applied in supporting those tasks that creatives deem less creative and can be a useful time saver when a creative can focus on making changes to their works in progress based on client, team, or user feedback. The illusion of generative AI being creative and that creativity itself can be correlated to intelligence is a generalization that is not supported by evidence-backed research nor by the pragmatic integration of AI across creative industries. What may mute debates of an AI taking over the job of an artist, for example, is that just as generative AI are programmed to seek patterns and generate content from those patterns, many artists look for patterns in order to break them, to move beyond them to create something that has not been created before. The ground-breaking part of AI as a technology is not in the code itself, but lies in the imagination of the creative that guides it to support unique acts of creation.

技术很好……人工智能很好

Tech Is Good… AI Is Good

对于一些创意人员来说，生成式人工智能是一种更有用的工具，如果公平地使用它来转换、重新语境化和融合他们自己的想法、文字和图像，以通过原型支持和激发他们的创意过程。对于其他人来说，它们的存在本身就被批评为鲁莽、危险，并且可能导致一些不可调和的困境长期存在。还有一些人拥抱使用生成人工智能作为抗议场所，以批评他们用来交流艺术的东西。致力于展示法学硕士严肃和愚蠢的可供性的网站不断涌现。生成式人工智能可以用来故意展示其无法陈述事实或不押韵诗歌，无法创造出太多的手，无法生成能够破坏模式化、公式化音乐的音乐作品，而这些音乐一直被创造出来并被宣传为世界上最好的音乐。如果公平、有意且适当地应用，生成式人工智能可以在设计不遵循规范的想法原型时激发您的创造性，以支持您自己的创作过程。就像约翰·格林·钱德勒 (John Greene Chandler) 的《小鸡》(Chicken Little) 中一样，尽管反对的呼声值得关注，但天并没有真正塌下来（图1-3）。在这个计算能力强大的时代，我们可以体会到生成式人工智能可以为我们提供的好处，并且许多人将并且已经出于各种不同的原因使用它们，从而造福于其他人类。

人工智能生成的鸡的照片。左边的鸡有一只眼睛，中间的鸡头上有两个王冠，右边的鸡身材矮小，有一个长喙。

A photograph of the A I generated chickens. The left chicken has one eye, the middle chicken has two crowns on his head, and the right chicken is short and has a long beak.

因此，让我们将其付诸实践，通过重塑钱德勒在人工智能吸引越来越多人类注意力的历史时期的著名故事来纪念他。

So let’s put that to practice, honoring Chandler by reinventing his well-known story for a time in history where AI is captivating the attention of an increasing number of humans.

作者：给我一个小鸡故事的版本：“天塌下来了。天塌下来了”，但使用人工智能。

Author: Give me a version of Chicken Little’s story with “The sky is falling. The sky is falling” but use AI.

LLM ：从前，在一个小村庄里，住着一只名叫小鸡的鸡。有一天，利特尔在地上啄食时，听到一种奇怪的声音。突然，一架人工智能气象无人机飞过头顶，投射出黑暗、暴风雨天空的全息图像。无人机宣布：“天塌下来了！天塌下来了！ （图 1-4）”

LLM: Once upon a time, in a small village, there lived a chicken named Little. One day, while pecking at the ground for food, Little heard a strange noise. Suddenly, an AI-powered weather drone flew overhead and projected a holographic image of a dark, stormy sky. The drone announced, “The sky is falling! The sky is falling! (Figure 1-4)”

一张人工智能生成的谷仓夜间照片。谷仓周围都是母鸡。两架无人机在谷仓上空盘旋。背景中有一片森林。

A photograph of an A I generated barn at night. The barn is surrounded by hens. Two drones are hovering above the barn. There is a forest in the background.

小小很害怕，跑去告诉村里的其他动物。他们都嘲笑她，认为这只是她又一个疯狂的想法。但利特尔决心警告他们危险并着手寻找解决方案。

Little was terrified and ran to tell the other animals in the village. They all scoffed at her, thinking it was just another one of her wild ideas. But Little was determined to warn them of the danger and set out to find a solution.

她遇到了一群正在研究人工智能天气预报系统的科学家（图 1-5）。他们向她解释说，人工智能检测到风暴即将来临，村庄需要为此做好准备。小小闻言松了口气，跑回村里告诉其他人这个好消息。

She came across a group of scientists who were working on an AI-powered weather forecasting system (Figure 1-5). They explained to her that the AI had detected an approaching storm and that the village needed to prepare for it. Little was relieved to hear this and ran back to the village to tell the others the good news.

一张人工智能生成的卡通照片，描绘了 5 只狐狸在笔记本电脑上工作。其中一只狐狸正在调整笔记本电脑顶部的天线。狐狸有双重色调。

A photograph of an A I generated cartoon depicting 5 foxes working on a laptop. One of the foxes is adjusting the antenna on the top of the laptop. The foxes are in double shades.

动物们很快开始为风暴做准备，建造庇护所并收集物资（图 1-6）。暴风雨袭来，但多亏了人工智能的警告和利特尔的决心，村庄才得以平安度过暴风雨。

A photograph of the A I generated chickens. The chicken on the left is short and has large eyes, the chicken in the middle has large cheeks, and the chicken on the right is grey and short. A bolt of lightning strikes in the distance.

A photograph of the A I generated chickens. The chicken on the left is short and has large eyes, the chicken in the middle has large cheeks, and the chicken on the right is grey and short. A bolt of lightning strikes in the distance.

从那天起，利特尔因其敏捷的思维和愿意相信人工智能的力量而被称为村里的英雄。这个村庄学会了信任人工智能天气预报系统，它帮助他们免受未来风暴的影响。

From that day on, Little was known as the village’s hero for her quick thinking and her willingness to believe in the power of AI. And the village learned to trust in the AI-powered weather forecasting system, and it helped them keep safe from future storms.

结束。

The end.

就讲故事而言，法学硕士提供的一个有趣的转折需要精心设计，但种子已经存在。我们有一架人工智能驱动的无人机，它似乎在自行行动并引起恐慌，后来才发现它是由科学家操作的，以警告小鸡和她的朋友前方的危险。人工智能显示了人类（或在本例中为狐狸）和人工智能内在的相互依赖性。对于那些熟悉《小鸡小鸡》故事的人来说，生成的版本在深度、寓言或丰富的人物塑造方面确实与原作相差甚远。然而，它可能成为新故事的种子。生成的故事还揭示了法学硕士必须包含在故事中的意外吹嘘。这种哗众取宠的行为也表明了绿野仙踪在幕后的表现，以及他们需要吹嘘他们的弗兰肯蹒跚学步的人。

An interesting twist offered up by an LLM needs crafting as far as storytelling goes, but the seeds are there. We have an AI-powered drone seemingly acting on its own and creating panic only to find out later that it is being operated by scientists to warn Chicken Little and her friends of the danger ahead. The AI shows the intrinsic interdependency of human (or in this case fox) and AI. For those familiar with the story of Chicken Little, the generated version really does not even come close to the original in depth, allegory, or rich characterization. It might, however, become the seed for a new story. The generated story also reveals the unexpected bragging that the LLM had to include as part of the story. That grandstanding is also an indicator of the Wizard of Oz behind the curtain and their need to boast about their Frankentoddler.

任何类型的生成故事都可以由任何创意人员策划，并且当不将其视为最终产品而是将其视为创意和关键流程的一部分时，可以更合乎道德地使用。各种各样的原型都成为可能，可以激发各种创作者的灵感。使用小鸡故事中的情节点和角色生成有关人工智能的故事的活动向我们展示了一个需要完善的原型。显而易见，生成式人工智能除了精炼人工智能生成的内容之外，还需要人类参与其中才能生成任何东西。每一个狭隘的人工智能都需要人类参与其中。需要引导他们有意图和目的。他们无法独自做到这一点。

Generated stories of all types can be curated by any creative and can be used more ethically when thought of not as an end product, but as part of your creative and critical process. All manner of prototypes are made possible that can inspire creators of all kinds. The activity of generating a story about AI using plot points and characters that were part of the Chicken Little story showed us a prototype that needs refinement. Stating the obvious, generative AI needs the human in the equation in order to generate anything, in addition to refining content that an AI generates. Every narrow AI requires the human in the equation. They need to be guided to have intention and purpose. They are unable to do that on their own.

调和人工智能的炒作与诽谤

Reconciling the Hype and the Vilification of AI

作为不断寻找新工具的创意人员，在使用生成式人工智能之前协调其使用非常重要。人工智能并不是第一个引起我们广泛关注的技术。我们对任何技术的正面和负面炒作的反应都是一种历史模式，作者兰登·温纳 (Langdon Winner) 在他1988 年出版的《鲸鱼和反应堆》一书中很好地阐述了这一点。对于温纳和他的书出版前后的许多思想家来说，对于正在采用的任何技术的社会、经济和政治层面，保持清醒和批判性的积极态度至关重要。人类与可编程机器的互动不断引发媒体对其利弊进行权衡，通过夸张夸张、不合理的恐惧和消极情绪夸大人工智能（AI）来争夺人们的注意力。您将阅读或聆听并观察正面和负面的头条新闻，包括那些基于网络的法学硕士，它们可以生成任何内容，从一篇关于艾伦图灵的影响的文章（省略了琼克拉克）到对一位作者写一本关于生成人工智能的书的极其不准确的描述在图书馆里。受欢迎的作家、思想家、政治家、科学家、艺术家和活动家进一步放大了观点。接下来引人注目的头条新闻是使用应用监督和强化学习方法的法学硕士生成的。作为您自己的创意过程的一部分，请让法学硕士研究人工智能的优点和缺点。您会发现，负面和正面的标题在被视为真相陈述之前都需要严格的关注和研究。人类生成的头条新闻也是如此。使用生成式人工智能的决定需要您自己的洞察力。

As a creative on the constant lookout for new tools, it’s important to reconcile your use of generative AI prior to using it. AI is not the first technology to gather our critical attention. Our reactions to the positive and negative hype of any technology are a historical pattern that was well articulated by author Langdon Winner in his book The Whale and the Reactor, published in 1988. For Winner and many thinkers before and after his book was published, it is crucial to be awake and critically active regarding the social, economic, and political dimensions of any technology that is in the process of being adopted. Human interactions with programmable machines have persistently triggered media outlets to weigh in on their pros and cons, to compete for your attention by dramatizing artificial intelligence (AI) with a mix of exaggerated hyperbole and unreasonable fear and negativity. You’ll read or listen to and observe positive and negative headlines including those web-based LLMs that can generate anything from an essay on the influence of Alan Turing that omits Joan Clarke to a ridiculously inaccurate depiction of an author writing a book on generative AI in a library. Opinions are further amplified by popular authors, thinkers, politicians, scientists, artists, and activists. The attention-grabbing headlines that follow were generated using an LLM that applied supervised and reinforcement learning methods. As part of your own creative process, task an LLM to generate the pros and cons of AI. You will see that both negative and positive headlines require critical attention and research prior to being taken for truth statements. The same can be said of human-generated headlines. The decision to use generative AI requires your own discernment.

天网蓬勃发展时可尝试的创意活动

Creative Activities to Try While Skynet Thrives

原子弹、人类末日以及对人工智能的恐惧都是用来谈论人工智能危险的隐喻。它们也可以成为创意温室。您可以使用这些故事、隐喻和警示故事作为您自己的科幻小说、歌曲、设计、漫画、网站、TikTok 短片、电视剧或博客文章的起点。

The atom bomb, the end of people, and being frightened with AI are all metaphors that are used to talk about the dangers of AI. They can also be creative hothouses. You can use these stories, metaphors, and cautionary tales as jumping-off points for your own science fiction, song, designs, manga, website, TikTok short, TV series, or blog post.

还有更多的艺术家将他们所处时代的逆境和挑战转化为强大的创意表达，这表明艺术和创意行为可以成为处理和交流困难经历以及批评既定社会实践和行为的深刻工具。所有学科的创意人员已经开始利用生成式人工智能作为社会批评的焦点。 Trevor Paglen 以人工智能和监控领域的工作而闻名，他利用人工智能和机器学习创建了一系列名为“机器视觉”的图像。该项目旨在批评和提出有关人工智能系统如何“训练”以观察和解释世界的问题。 Holly Herndon、Taryn Southern、Dadabots、Ash Koosha 等音乐艺术家强调了人工智能作为创意工具的潜力，并提出了关于作者身份、创造力以及日益自动化的世界中音乐的未来的重要问题。

There are many more artists who have transformed their adversity and the challenges of their time into powerful creative expressions, showing that art and creative acts can be a profound vehicle for processing and communicating difficult experiences and critiquing established social practices and behaviors. Creatives of all disciplines are already engaged in leveraging generative AI as a place of focus for social critique. Trevor Paglen, known for his work with AI and surveillance, used AI and machine learning to create a series of images titled “Machine Visions.” The project was designed to critique and raise questions about how AI systems are “trained” to see and interpret the world. Musical artists like Holly Herndon, Taryn Southern, Dadabots, Ash Koosha, and others highlight the potential of AI as a creative tool and also raise important questions about authorship, creativity, and the future of music in an increasingly automated world.

五个为什么

Five Why’s

您还可以应用创意人士使用的设计工具来找到问题的根源。 “五个为什么”练习不断地询问团队试图识别的特定问题的原因。询问我们的法学硕士是否熟悉设计思维活动会得到以下答复。

You can also apply a design tool that creatives have used to get to the root of a problem. The Five Why’s exercise continuously asks why to a particular problem that a team is trying to identify. Asking our LLM if it is familiar with the design thinking activity generates the following response.

LLM：“五个为什么”是一种解决问题的工具，可以鼓励人们更深入地了解当前的问题。最初由丰田佐吉开发，并在丰田汽车公司制造方法的演变过程中使用，现在已应用于各个行业。这个过程涉及反复询问“为什么？”揭开问题的各个层面并找到根本原因。识别出最初的问题后，您会问：“为什么这是一个问题？”这个问题的答案就成为下一个“为什么？”的基础。问题等等。这一直持续到你问“为什么？”五次或直到弄清楚问题的根本原因是什么。

LLM: The Five Why’s is a problem-solving tool that encourages a deeper understanding of the issue at hand. Originally developed by Sakichi Toyoda and used within the Toyota Motor Corporation during the evolution of its manufacturing methodologies, it’s now used in various industries. The process involves repeatedly asking “Why?” to peel back the layers of an issue and get to the root cause. After an initial problem is identified, you ask, “Why is this a problem?” The answer to that question then becomes the basis for the next “Why?” question and so forth. This continues until you’ve asked “Why?” five times or until it’s clear what the root cause of the problem is.

如果担心人工智能在各方面变得像人类并最终取代人类，你可以质疑这个假设本身。从“为什么”这个词开始，生成你自己的线索，或者更好的是，构建五个为什么的活动，让法学硕士来回答。

In the case of the fear of AI becoming human-like in every way and eventually dispensing of humans, you can interrogate the assumption itself. Start with the word “Why” and generate your own thread or, better yet, structure the Five Why’s activity for an LLM to answer.

作者：将五个为什么应用于以下陈述：人类担心人工智能会变得有意识并且不再需要它们。

Author: Apply the Five Why’s to the following statement: humans fear that AI will become conscious and no longer need them.

LLM：当然，让我们将五个为什么应用于该陈述。

LLM: Absolutely, let’s apply the Five Why’s to that statement.

为什么人类担心人工智能会变得有意识而不再需要它们？

Why do humans fear that AI will become conscious and no longer need them?

因为他们担心失去对人工智能的控制，这可能会导致潜在的伤害或人类角色的取代。

Because they worry about losing control over AI, which might lead to potential harm or displacement of human roles.

为什么人类担心失去对人工智能的控制？

Why do humans worry about losing control over AI?

因为人工智能系统如果有意识和自主，可能会做出不符合人类利益或道德标准的决策。

Because AI systems, if conscious and autonomous, might make decisions that are not aligned with human interests or ethical standards.

为什么人工智能系统会做出不符合人类利益或道德标准的决策？

Why might AI systems make decisions that are not aligned with human interests or ethical standards?

因为人工智能，特别是如果它实现了意识，会根据其编程而行动，而不一定是人类的道德或情感。如果它没有经过适当的编程来考虑这些因素，它可能会做出对人类有害或令人不安的决定。

Because AI, especially if it achieved consciousness, would act based on its programming and not necessarily human ethics or emotions. If it’s not properly programmed to consider these factors, it may make decisions that are harmful or unsettling to humans.

为什么人工智能不能被适当地编程来考虑人类的道德和情感？

Why may AI not be properly programmed to consider human ethics and emotions?

因为即使开发人员努力将道德考虑嵌入到人工智能中，捕捉人类道德和情感的复杂性和细微差别仍然极具挑战性。此外，全球对某些道德标准缺乏共识。

Because even though developers strive to embed ethical considerations into AI, it’s extremely challenging to capture the complexity and nuances of human ethics and emotions. Also, there’s a lack of global consensus on certain ethical standards.

为什么在人工智能编程中捕捉人类道德和情感的复杂性和细微差别具有挑战性？

Why is it challenging to capture the complexity and nuances of human ethics and emotions in AI programming?

因为人类的道德和情感非常复杂、依赖于具体情况，而且往往是主观的。它们已经进化了数千年，并受到文化、个人和情境因素的影响，这些因素很难全面编码到人工智能系统中。

Because human ethics and emotions are deeply complex, context-dependent, and often subjective. They have evolved over thousands of years and are influenced by cultural, personal, and situational factors that are difficult to comprehensively codify into an AI system.

您可以应用自己的提示或创建“人类担心人工智能会变得有意识并且不再需要它们”的变体，或者简单地重新生成提示并查看人工智能是否给您另一个响应。

You can apply your own prompt or create a variation of “humans fear that AI will become conscious and no longer need them,” or simply regenerate the prompt and see if the AI gives you another response.

当你应用“五个为什么”来促使法学硕士做出回应时，即使是对 AGI 毫无根据的恐惧也可以转变为创造性的目的。在此过程中，您可能会发现“五个为什么”与生成式人工智能相结合可以产生一些您可以询问的有趣的根本原因问题，并且该活动可以应用于其他设计问题。

As you apply the Five Why’s to prompt a response from an LLM, even an unfounded fear of AGI can be turned around for creative purposes. In doing so you might discover that the Five Why’s combined with a generative AI can yield several interesting root cause problems you can interrogate, and the activity can be applied to other design problems.

作为创意人员，面对生成式人工智能的相互冲突的情绪反应是很重要的。艺术家们坚持批判性地对待技术，并将这种批评融入到他们的创造性表达中。作为一个有创造力的人，你不会遵循危言耸听或夸张的模式，而是会从发展自己使用生成人工智能的理由中受益。您可能会认为，过去两个世纪艺术家和设计师的大部分目的都是为了打破惯例、破坏社会礼仪和拆除既定形式。虽然有些创意人员是为了钱，但你可能不会仅仅因为它们被证明是成功的而对重复以前的模式感兴趣。想想那些打破了在自己的学科（如视频游戏、音乐或视觉艺术、舞蹈、戏剧或电影）中展示自己创作的模式的创意人员。他们的共同点是反复尝试破坏观众的期望。您如何整合生成式人工智能来颠覆您自己既定的创造模式？生成式人工智能如何支持您自己的创作过程的发展？为什么您觉得有必要这样做？

As creatives, the conflicting emotional reactions to generative AI are important to face. Artists have persistently taken on technology critically and integrated that criticism into their creative expression. Rather than following the pattern of fearmongering or hyperbole, as a creative person you will benefit from developing your own rationale for using generative AI. You might consider that much of the purpose of artists and designers in the last two centuries has been about breaking convention, disrupting social etiquettes, and dismantling established forms. While some creatives are in it for the money, you may not be as interested in repeating patterns of what came before simply because they proved successful. Consider those creatives that have broken patterns of presenting their creations within their own disciplines like video games, music or visual art, dance, theater, or film. What they share in common are their iterative attempts to disrupt an audience’s expectations. How might you integrate generative AI to disrupt your own established patterns of creation? How might generative AI support your own creative process to evolve? Why might you feel a need to do so?

无论头条新闻具有挑衅性，也无论其野火蔓延的病毒性质如何，法学硕士都要求我们就生成人工智能进行一场非常真实的对话，以某种方式形成观点。这些观点可以是通过您自己的研究获得的丰富信息，与您通常阅读或聆听的人一致，也可以是基于您的直觉、生活经验、您看待世界的方式以及技术在其中的作用的直觉反应。无论生成式人工智能支持、中断或拆除哪些人类活动，表达你对生成式人工智能观点的批评声音都很重要。正如无数其他创新技术一样，生成式人工智能也是如此。选择你自己的冒险，了解包括你自己在内的批评声音，并为了他人的利益而改变你对技术的使用（图1-7）。

A photograph of an A I generated image depicts robotic hands grasping a hollow globe of Earth. A small robot is placed on the globe. Tiny planets are scattered around the globe.

A photograph of an A I generated image depicts robotic hands grasping a hollow globe of Earth. A small robot is placed on the globe. Tiny planets are scattered around the globe.

致谢

Acknowledgments

参考

References

巴特勒，S.（1863）。 “机器中的达尔文”，205。

Butler, S. (1863). “Darwin Among the Machines,” 205.

钱德勒，JG (1855)。小鸡的非凡故事。 WJ 雷诺兹公司，康希尔 24 号。

Chandler, J. G. (1855). Remarkable Story of Chicken Little. W. J. Reynolds & Company, 24 Cornhill.

洛夫莱斯，AA (1843)。译者注释（对 LF Menabrea 的“查尔斯·巴贝奇发明的分析机的草图”的翻译的附录）。科学回忆录，选自《外国科学院学报》和《外国期刊》，3, 691–731。

Lovelace, A. A. (1843). Notes by the translator (addenda to her translation of LF Menabrea’s “Sketch of the Analytical Engine Invented by Charles Babbage”). Scientific Memoirs, Selected from the Transactions of Foreign Academies of Science and Learned Societies, and from Foreign Journals, 3, 691–731.

温纳，L. (2010)。鲸鱼和反应堆：在高科技时代寻找极限。芝加哥大学出版社。

Winner, L. (2010). The Whale and the Reactor: A Search for Limits in an Age of High Technology. University of Chicago Press.

智能机器

Intelligent Machines

如果你进行创作，那么在你的创作过程中的某个时刻，你将不可避免地与某种技术进行交互。技术来自于我们的想象，而不是相反。技术不断激发艺术创作、设计和人类世界体验的新创意方法。支持人类创造力的技术清单非常详尽。对我来说幸运的是，我有一名法学硕士可供我使用，他可以提供您可以进行事实核查的粗略历史。

If you create, then at some point in your creative process, you will inevitably interact with some kind of technology. Technologies have come from our imagination, not the other way around. Technology has persistently incited new creative approaches to artmaking, design, and the human experiences of the world. The list of technologies that have been imagined to support human creativity is exhaustive. Lucky for me, I have an LLM at my disposal who can provide a rough history that you can fact-check.

LLM：以下是一些历史亮点。

LLM: Here are some historical highlights.

石器 （约 260 万年前）：人类创造力和表达的最早证据来自史前时代，当时我们的祖先制作石器。这些工具不仅用于生存任务，还用于在洞穴墙壁和物体上雕刻和创造早期艺术形式。

Stone Tools (~2.6 million years ago): The earliest evidence of human creativity and expression comes from prehistoric times when our ancestors crafted stone tools. These tools were not only used for survival tasks but also for carving and creating early art forms on cave walls and on objects.

涂料 （约 4 万年前）：涂料的出现使得洞穴壁画中的描绘变得更加复杂和生动。它开辟了一种全新的视觉叙事方式。

Paint (~40,000 years ago): The creation of paint allowed for more complex and vibrant depictions in cave paintings. It opened up a whole new way of visual storytelling.

书写工具 （约公元前 4000 年）：手写笔和早期墨水的发明使古代文明能够在泥板和纸莎草上记录信息并创作艺术品。

Writing Instruments (around 4000 BC): The invention of stylus and early forms of ink allowed ancient civilizations to record information and create artwork on clay tablets and papyrus.

纸 （大约公元 100 年）：纸发明于中国，为艺术家提供了一种轻质、便携的表面，供艺术家绘画。这彻底改变了人们创作和分享艺术的方式。

Paper (around 100 AD): Invented in China, paper provided a lightweight, portable surface for artists to draw and paint on. This revolutionized the way people could create and share art.

印刷机 （1440）：约翰内斯·古腾堡发明的印刷机使得书籍和图像的大规模生产成为可能。这不仅使信息民主化，而且影响了艺术风格，导致了蚀刻和雕刻等技术的出现。

Printing Press (1440): Johannes Gutenberg’s invention of the printing press allowed for the mass production of books and images. This not only democratized information but also influenced artistic styles, leading to techniques like etching and engraving.

相机 （十九世纪）：摄影的发明使艺术家能够以一种新的、创新的方式捕捉现实。它还影响了绘画等许多其他艺术形式，导致了现实主义等运动，并后来激发了抽象表现主义等新的艺术形式。

Camera (nineteenth century): The invention of photography allowed artists to capture reality in a new and innovative way. It also influenced many other art forms like painting, leading to movements such as realism and later inspiring new art forms like abstract expressionism.

合成颜料 （十九世纪中叶）：合成颜料的发展扩展了艺术家的调色板，超越了传统的天然颜料。这使得艺术作品更加充满活力和持久。

Synthetic Paint (mid-nineteenth century): The development of synthetic pigments expanded the artist’s color palette beyond the traditional, naturally derived pigments. This allowed for more vibrant and long-lasting artwork.

电影 （十九世纪末）：电影的发明为艺术家探索和表达创造力开辟了全新的媒介，从而导致了电影业的诞生。

Film (late nineteenth century): The invention of moving pictures opened up a completely new medium for artists to explore and express their creativity, leading to the birth of the film industry.

计算机 （二十世纪中叶）：计算机以及后来的图形软件为艺术家提供了创作数字艺术的新媒介。这扩大了创建和共享内容的可能性。

Computer (mid-twentieth century): Computers, and later graphics software, gave artists a new medium to create digital art. This expanded the possibilities of what could be created and shared.

互联网 （二十世纪末）：互联网彻底改变了艺术的共享和消费。它为艺术家提供了一个与世界分享作品的平台，并开辟了新的艺术形式，如网页设计和数字动画。

Internet (late twentieth century): The Internet revolutionized the sharing and consumption of art. It provided a platform for artists to share their work with the world and also opened up new forms of art, like web design and digital animation.

3D打印 （二十世纪初）：3D打印使艺术家能够用各种材料创作复杂的三维作品，为雕塑和装置艺术开辟了新的机遇。

3D Printing (early twenty-first century): 3D printing has allowed artists to create complex, three-dimensional works in a variety of materials, opening up new opportunities for sculpture and installation art.

你可以看到法学硕士关于技术构成的想法是广泛的。在技​​术的众多定义中，我最喜欢的是科幻作家兼幽默家道格拉斯·亚当斯的定义，他说过“技术”一词是用来描述尚未完全发挥作用的事物的。人类长期以来一直痴迷于共同构建可以支持创造行为的不完美技术。

You can see that the LLM’s ideas of what constitutes technology are broad. Amid the numerous definitions of technology, my favorite is science fiction writer and humorist Douglas Adams’s, who said something like technology being a word that describes something that doesn’t quite work yet. Humans have long been obsessed with co-constructing imperfect technology that can support acts of creation.

与生成式人工智能互动的原因之一是从中汲取灵感。在创作过程中，创意人员一直在相互合作。当周围没有人时，您可以使用人工智能来快速测试想法。这些可能会给你关于下一步该做什么或者一个完全充实的视觉想法可能是什么样子的想法。法学硕士甚至可能会就您的想法提供反馈。如果您对想要创造的东西有一个令人惊叹的想法，并且您想知道在您之前是否有人已经做过，那么情况尤其如此。有时我们没有语言来表达我们想要发明什么，因此如果不投入大量时间和精力，在互联网上进行搜索可能不会得到结果。虽然这类似于法学硕士回应您的查询的方式，但您可能会更快地获得知识领域的广泛概念，或者人工智能可能会指出您在其他地方找不到的直接来源，您可以在网上进一步研究。我自己对使用智能机器进行创作的兴趣和投资受到了我童年时与自动钢琴的互动的影响，那么为什么不促使法学硕士讲述自己的起源故事呢？创造性地这样做可能会产生成果，并展示如何从人工智能中激发想法，以推动进一步的创造性研究。

One reason to interact with a generative AI is to bounce ideas off it. Creatives do this all the time with one another during the creation process. When there’s no one around, you can access an AI to test out ideas rapidly. These might give you ideas as to what to do next or what a fully fleshed-out visual idea might look like. An LLM might even give you feedback on an idea that you have. This is especially the case if you have an amazing idea for something you want to create and you want to know if it has been done by others before you. Sometimes we don’t have the language to articulate what it is we want to invent, so conducting a search on the Internet might not get you results without a lot of time and effort put into it. While this is similar to how an LLM might respond to your queries, you will likely receive a broad idea of a knowledge area faster, or the AI might point to a direct source you couldn’t find elsewhere that you can further research online. My own interest and investment in using intelligent machines to create was influenced by interactions with a player piano in my childhood, so why not prompt an LLM to tell its own origin story? Doing so creatively may yield results and demonstrate how you can bounce ideas off an AI to propel further creative research.

作者：从历史角度讲，您会从哪里开始讲述机器智能的故事？为什么？

Author: Where would you start historically in telling the story of machine intelligence and why?

LLM：机器智能的故事可以追溯到古代，当时哲学家和发明家第一次开始考虑创造可以模仿人类智能的机器。然而，随着计算机的出现和 人工智能（AI） 作为一门学科的发展，对机器智能作为一个研究领域的更全面的理解直到二十世纪才开始出现。标志着现代人工智能开始的一个重大事件是 1956 年的达特茅斯会议，该会议被广泛认为是人工智能这一研究领域的诞生地。这次会议聚集了一群计算机科学家、数学家和工程师，他们都有共同的兴趣，即探索计算机执行通常需要人类智能的任务的潜力，例如识别模式、从经验中学习和解决问题。

LLM: The story of machine intelligence can be traced back to ancient times, when philosophers and inventors first began contemplating the idea of creating machines that could mimic human intelligence. However, a more comprehensive understanding of machine intelligence as a field of study only began to emerge in the twentieth century, with the advent of computers and the development of artificial intelligence (AI) as a discipline. One significant event that marked the beginning of modern AI was the Dartmouth Conference in 1956, which is widely considered as the birthplace of AI as a field of study. The conference brought together a group of computer scientists, mathematicians, and engineers who shared a common interest in exploring the potential of computers to perform tasks that typically require human intelligence, such as recognizing patterns, learning from experience, and solving problems.

当你从法学硕士那里得到想法时，你会参与一个迭代过程，随着你与人工智能的对话关系的发展，这个过程可以激发你变得更加具体。这个过程被称为连锁提示。在前面的例子中，我获得了两个新的见解。第一个是发明家想要创造模仿人类智能的机器的想法，这让我很好奇哪些发明家、什么机器以及人类痴迷于这样做的时间有多早。第二个是进一步研究人工智能一词何时出现的良好起点，它的定义是什么，谁出现在历史的那个时刻，以及在那段时间可能发生了哪些其他发展。前面示例中提示 AI 的想法也可以重写为文本图像生成 AI 的输入，以了解 AI 在搜索自己的起源故事时如何描绘自己（图2-1）。

When you bounce ideas off an LLM, you engage in an iterative process that can inspire you to get more specific as your conversational relationship with an AI develops. This process is referred to as chain prompting. In the preceding example, I gained two new insights. The first was the idea of inventors wanting to create machines mimicking human intelligence that makes me curious about which inventors, what machines, and how far back in time humans have been obsessed with doing so. The second was a good starting place to further research when the term artificial intelligence came to be, what the definitions of it were, who was present in that moment in history, and what other developments might have been occurring during that time. The idea of prompting an AI in the preceding example can also be rewritten as input for a text-image generative AI to see how an AI might depict itself as it searches for its own origin stories (Figure 2-1).

A photograph of a robot looking at art in a museum.

A photograph of a robot looking at art in a museum.

自动机

Automata

在古希腊，“自动机”一词的意思类似于“按照自己的意愿行事”。自动机的记录记录揭示了两个重要的考虑因素：第一是它们通常被构造为模仿需要一定程度智能的人类活动，第二，它们也被用作支持人类活动的功能工具，例如时钟，或者它们代表可以大规模建造的小型化原型。

In ancient Greece the word automata meant something akin to “acting of one’s own will.” Documented accounts of automata reveal two important considerations: the first is that they often were constructed to mimic human activities that required a degree of intelligence, and second, they were also used as functional tools to support human activities such as a clock, or they represented miniaturized prototypes that could be built on a larger scale.

通常建造微型自动机是为了演示事物如何运作并展示一些科学原理。海伦 (Heron) 的翻译文本对此进行了解释，他是一位才华横溢的数学家、物理学家和工程师，生活在公元 10-70 年左右。他的三部著作《气动》、《机械》和《自动机》提供了存在数百种能够自动运动的各种类型机器的证据。他的叙述表明，古希腊人正在将自动机（机械装置）的概念融入他们的日常生活中。海伦的发明包括自动门、一只唱歌的鸟和一个完整的自动化木偶剧院，能够使用绳索、结和简单的机器播放十分钟的戏剧。

Miniature automata were often constructed in order to demonstrate how things worked and to show some scientific principles. This is accounted for in the translated texts of Heron who was a talented mathematician, physicist, and engineer and lived around 10–70 AD. His three written works, “Pneumatica,” “Mechanica,” and “Automata,” provide evidence of the existence of hundreds of various types of machines capable of automated movement. His accounts show that the ancient Greeks were incorporating the concept of automata (mechanical devices) into their daily lives. Heron’s inventions included automated doors, a singing bird, and an entire automated puppet theater capable of playing a ten-minute drama using ropes, knots, and simple machines.

法学硕士：不要忘记，以机器人为特色的艺术表现和戏剧作品可以追溯到公元前三世纪左右的汉朝古代中国。在此期间，一个令人印象深刻的机械管弦乐队被制作出来，还有各种各样的机械化玩具，如空中自动机、鸽子和鱼、天体和神话龙的机械代表，以及自动操作的杯子服务器。这些机械奇迹主要由液压机构驱动，是工程师和工匠专门为娱乐皇帝而设计的，而他们的身份在很大程度上已经被遗忘。

LLM: Don’t forget that artistic representations and theatrical productions featuring robotics can be traced back to ancient China during the Han dynasty, around the third century BC. During this time, an impressive mechanical orchestra was crafted, along with an assortment of mechanized playthings such as airborne automatons, mechanical representations of doves and fish, celestial beings and mythical dragons, and self-operating cup servers. These mechanical marvels, predominantly driven by hydraulic mechanisms, were specifically designed to entertain emperors by engineers and craftsmen, whose identities have largely faded into oblivion.

由法学硕士引发的更多研究最终可能会导致唐朝的《水利诸侯书》。中国的机械奇迹通常由水、重力或其他简单的机械方法提供动力。其中许多设备是出于实际原因而创建的，例如测量时间或检测自然现象（图2-2）。不过，它们也有娱乐用途，其中机械娃娃和木偶戏颇受欢迎。

A photograph of a chariot. It has a transparent body resting on two large wheels.

A photograph of a chariot. It has a transparent body resting on two large wheels.

中国的发明可分为以下几类：

• 科学仪器：张衡的地动仪或苏松的天文钟楼，用于测量和记录科学数据。

• Scientific Instruments: Zhang Heng’s seismoscope or Su Song’s astronomical clock tower used to measure and record scientific data.

• 计时装置：许多自动机都是精致的时钟，使用机械运动来指示时间。

• Timekeeping Devices: Many automata were elaborate clocks, using mechanical movements to indicate time.

• 机械玩具和木偶：这些用于娱乐目的，包含可以移动或执行动作的人物。

• Mechanical Toys and Puppets: These were used for entertainment purposes incorporating figures that would move or perform actions.

• 乐器：朱载堉的自动笛子是为了通过复杂的机械手段演奏音乐而设计的。

• Musical Instruments: Zhu Zaiyu’s automatic flute player was designed to play music through complex mechanical means.

• 水利发明：在《水利优雅之书》中，这些发明使用水来驱动机械人物或执行其他任务。

• Hydraulic Inventions: In the Book of Hydraulic Elegancies, these used water to animate mechanical figures or perform other tasks.

古代音乐机器人乐队

Ancient Musical Robot Bands

预编程智能机器已经存在很长时间了。在大多数情况下，它们仅限于创建者设计和构建的特定操作。将它们视为预先编程的移动机器人，它们执行重复的特定机械运动。 Al-Jazari于 1206 年撰写的《巧妙机械装置知识之书》详细介绍了 50 种机械装置并提供了其构造说明。书中描述的最著名的设备之一是音乐机器人乐队，它被认为是最早的可编程自动机之一（图2-3）。 Al-Jazari 以创造机械设备的独创性而闻名，他设计了一种船形音乐自动机，其中有四名自动音乐家。这个音乐机器人被用来在皇家酒会上招待客人。

Pre-programmed intelligent machines have been around for a long time. For the most part, they were limited to specific actions that a creator designed and built. Think of these as pre-programmed moving robots that perform specific mechanical motions that repeat. Al-Jazari’s The Book of Knowledge of Ingenious Mechanical Devices, which he wrote in 1206, details 50 mechanical devices and provides instructions on their construction. One of the most notable devices described in the book is the musical robot band, considered to be one of the first programmable automata (Figure 2-3). Al-Jazari, known for his ingenuity in creating mechanical devices, designed a musical automaton in the form of a boat featuring four automated musicians. This musical robot was used to entertain guests at royal drinking parties.

A photograph of a painting of a royal drinking party with a mechanism below.

A photograph of a painting of a royal drinking party with a mechanism below.

谢菲尔德大学人工智能、机器人学和公众参与教授 Noel Sharkey表示，Al-Jazari 的音乐机器人乐队是可编程自动机的早期实例（图2-4）。夏基还尝试重新创建该机制，其特点是由凸轮控制的鼓机，凸轮触发打击乐器的杠杆。通过重新排列凸轮的位置可以改变鼓的图案。自动机不仅播放音乐，还在每次音乐表演中表演 50 多个面部和身体动作。

Professor of AI and robotics and public engagement at the University of Sheffield Noel Sharkey suggests that Al-Jazari’s musical robot band was an early instance of a programmable automaton (Figure 2-4). Sharkey has also attempted to recreate the mechanism, which features a drum machine controlled by cams that trigger levers for the percussion instruments. The drum patterns could be altered by rearranging the positions of the cams. The automata not only played music but also performed over 50 facial and body movements during each musical performance.

A photograph of an automated musical robot with a drum machine.

A photograph of an automated musical robot with a drum machine.

书写和绘画的自动机

Automata That Wrote and Drew

文本-​​文本和图像生成机器学习模型的古老祖先可以在瑞士发明家的机械魔法中编程的自动机中看到。这些圆滑而复杂的机械“魔鬼”在日益机械化的十八世纪欧洲风靡一时（图2-5）。

The ancient ancestor of text-text and image-generating machine learning models can be seen in the automata that were programmed in the mechanical wizardry of Swiss inventors. These slick and complicated mechanical “devils” became all the rage in a growing mechanistic eighteenth-century Europe (Figure 2-5).

A photograph of two robot boys on both sides of a female robot. Each is in a transparent box. The boys write and the female plays music.

A photograph of two robot boys on both sides of a female robot. Each is in a transparent box. The boys write and the female plays music.

1768 年至 1774 年间，皮埃尔·雅克德罗 (Pierre Jaquet-Droz)、他的儿子亨利·路易 (Henri-Louis) 和让·弗雷德里克·雷肖特 (Jean-Frédéric Leschot) 合作创造了三台非凡的自动机，包括“作家”（由 6000 个零件组成）、“音乐家”（由 2500 个零件组成）部分）和“绘图员”（2000 部分）。这些工程小奇迹以其复杂的机制吸引了欧洲、中国、印度和日本的观众。一些专家将这些设备视为计算机最早的例子。 “作家”是一个用羽毛笔和真正的墨水在纸上写字的机械男孩，他有一个制表符设置输入设备，可用作可编程存储器。它由 40 个摄像头驱动，充当其只读程序。这些作品是人类在解决机械问题方面最伟大的成就之一。跟随雅克德罗 (Jaquet-Droz) 的发条步伐，发明家亨利·梅亚德 (Henri Maillardet) 于 1805 年建造了一台弹簧驱动的自动机，可以用法语和英语绘制图像和创作诗歌。自动机的手部运动是通过位于底座轴上的一系列凸轮创建的，这些凸轮产生执行七个草图和随附文本所需的运动。该自动机被认为是当时所有自动机中拥有最广泛的基于凸轮的存储器。

Between 1768 and 1774, Pierre Jaquet-Droz, his son Henri-Louis, and Jean-Frédéric Leschot collaborated to create three remarkable automata, consisting of “The Writer” (composed of 6000 parts), “The Musician” (made up of 2500 parts), and “The Draughtsman” (2000 parts). These little marvels of engineering captivated audiences in Europe, China, India, and Japan with their intricate mechanisms. Some experts view these devices as the earliest examples of computers. “The Writer,” a mechanical boy who writes with a quill pen and real ink on paper, has a tab-setting input device that functions as a programmable memory. It is powered by 40 cams that act as its read-only program. The works are some of the greatest human achievements in mechanical problem-solving. Following in the clockwork pace of Jaquet-Droz, inventor Henri Maillardet constructed a spring-driven automaton in 1805 that could draw images and compose verses in both French and English. The automaton’s hand movements were created through a series of cams positioned on shafts at the base, which generated the necessary movement to execute seven sketches and accompanying text. This automaton is considered to have the most extensive cam-based memory of any automaton from that time.

当钢琴能弹钢琴时为什么还要弹钢琴？

Why Play the Piano When the Piano Can Play the Piano?

钢琴的发明比雅克德罗（Jaquet-Droz）早一点，值得指出，它是一项已经迭代了三个多世纪的技术的一个很好的例子。克里斯托福里的关键创新是一种允许用锤子敲击钢琴内部琴弦的机制。 “锤子动作”使演奏者能够根据施加在琴键上的力来控制音符的响度。这种表现力的品质使作曲家能够创作出具有表现力范围和动态控制的音乐，探索了音乐表现力的广阔新领域。钢琴是一项创新，引发了音乐创作的革命，钢琴成为十八、十九世纪欧洲新兴中产阶级的必备创新。人类不可避免地会进化出能够自动弹奏的钢琴。自二十世纪之交发明以来，自动钢琴作为一种看似智能且富有创造力的机器，不断发展，至今仍然在酒店大堂和餐厅中神奇地自行演奏，而不需要被注意到。

A bit earlier than Jaquet-Droz, it is worth pointing to the invention of the piano as a good example of a technology that has been iterated on for over three centuries. Cristofori’s key innovation was a mechanism that allowed for the strings inside the piano to be struck with a hammer. The “hammer action” enabled the player to control the loudness of the note based on the force applied to the keys. This expressive quality allowed composers to create music with an expressive range and dynamic control who explored a vast new spectrum of musical expression. The piano is an innovation that led to a revolution in music composition, and the piano became the must-have innovation for a rising European middle class in the eighteenth and nineteenth centuries. It was inevitable that humans would evolve a piano that could play itself. The evolution of the player piano since its turn-of-the-twentieth-century invention, as a seemingly intelligent and creative machine, still has sonic presence in hotel lobbies and restaurants magically playing on its own without the need to be noticed.

二十世纪之交获得专利的自动钢琴显示了人类对机器像人类一样演奏的渴望。我们还看到了具有可编程功能的机器。自动演奏钢琴由录制在穿孔纸或金属卷上的编程音乐组成。这些使音乐制作变得更加容易，任何人都可以使用踏板和控制杆控制节奏和其他效果。也被称为价格较高的再现钢琴，甚至可以模仿艺术家演奏的细微差别，要求用户通过重复踩下立式钢琴下方的脚踏板来简单地泵出音乐。最终，它们变成了电力驱动，最大限度地减少了用户的努力。在娱乐场所内放置的投币式钢琴中，我们不仅看到了自动点唱机和其他设备的出现，而且还看到了随着生成式人工智能的发展而不断发展的付费机器模式的支持者。自动钢琴是最早能够“存储”数据并播放数据的机器的例子。音符形式的数据被“加载”，然后可以通过一些人机交互来激活，以按预先编程的顺序弹奏钢琴上的特定琴键。文本音乐生成人工智能和自动演奏钢琴有很多共同点。这两个系统都依赖于某种形式的编码信息。在自动钢琴中，这是代表音符及其计时的穿孔卷轴。在文本到音乐的生成人工智能中，编码信息是可能描述音乐特征或可能根据特定规则或模式转换为音乐的文本输入。就像自动钢琴根据钢琴卷帘中的编码信息自动演奏音乐一样，文本到音乐人工智能根据对文本输入的理解生成并演奏一段音乐。自动钢琴通过机械地读取卷轴来生成音乐，而文本到音乐的人工智能则通过算法解释文本输入来生成音乐。

Player pianos patented close to the turn of the twentieth century show the human desire for machines to perform like a human, literally. We also see a machine with programmable capabilities. Player pianos consisted of programmed music recorded on perforated paper or metallic rolls. These made music production easier by allowing anyone the ability to control tempo and other effects with a treadle and levers. The pricier reproducing pianos as they were also called could even imitate the playing nuances of an artist tasking users to simply pump the music out by pressing on a foot pedal beneath the upright piano repetitively. Eventually they became electrically powered minimizing the effort of the user. In coin-operated pianos placed inside of entertainment venues, we not only see the advent of the jukebox and other devices, but we have a proponent for the pay-the-machine model that is evolving with the development of generative AI. Player pianos were examples of some of the first machines to be able to “store” data and play it back. That data in the form of musical notes was “loaded” and then could be activated by some human interaction to play specific keys on the piano in pre-programmed sequences. Text-music generative AI and the player piano share a lot in common. Both systems rely on some form of coded information. In a player piano, this is the perforated roll that represents musical notes and their timings. In a text-to-music generative AI, the coded information is the text input that might describe musical characteristics or might be transformed into music based on specific rules or patterns. Just like a player piano automatically performs music based on the coded information in the piano roll, a text-to-music AI generates and performs a piece of music based on its understanding of the text input. A player piano generates music by mechanically reading through a roll, and a text-to-music AI generates music by algorithmically interpreting the text input.

请注意，两者之间存在明显差异。关键在于，自动演奏钢琴并没有真正“编码”来生成像文本到音乐人工智能那样的音乐混搭，其结果可能从令人惊讶到绝对可怕。生成的流行音乐让音乐艺术家感到震惊，因为他们抱怨这听起来像他们，没有请求许可，而且他们基本上被排除在音乐创作过程之外。然而，回想一下，人工智能将根据其训练的数据生成内容。它将寻找模式，如果这些模式导致的内容在本例中听起来像原始内容，那么它的工作就完成了。对于音乐才华横溢的创意人士来说，重要的一课是认识人工智能正在生成哪些音乐模式。许多流行音乐听起来都一样，因为歌曲往往依赖于重复的基础和弦模式、节奏或别人音乐的采样块。每种音乐风格都有自己的模式，唱片公司依靠这些模式来继续销售该音乐。很少有人提出新的和弦、旋律和节奏模式。生成式人工智能可以为我们提供一首具有颠覆性的声音生成歌曲，这可能会激励我们打破惯例，并利用它生成的内容来推动创新的声音创作。

Mind you there are stark differences. The key one is that player pianos were not really “coded” to generate musical mashups like text-to-music AI can, whose results can range from astonishing to absolutely terrible. Generated pop music has alarmed musical artists because they complain that it sounds like them, that no permission was requested, and that they are basically left out of the musical creation process. Recall, however, that an AI will generate content based on the data it is trained on. It will look for patterns, and if those patterns lead to content that in this case sounds like the original, then its job is done. For musically talented creatives, the important lesson is to recognize what patterns in music an AI is generating. A lot of pop music sounds the same because songs tend to rely on repetitive underlying chordal patterns, rhythms, or sampled chunks of someone else’s music. Each style of music has its own patterns that record companies rely on in order to continue to sell that music. Rarely are new chordal, melodic, and rhythmic patterns proposed. Generative AI can offer us a disruptive sounding generated song that might inspire us to break convention and use what it generates to propel an innovative sonic creation.

创意人员的要点

Key Takeaways for Creatives

• 迭代开发：自动演奏钢琴随着时间的推移的发展展示了迭代和坚持的力量。自动钢琴的第一批型号通常非常基本，但通过不断的改进和创新，它们演变成更加复杂和功能强大的设备。应用于生成式人工智能，通过迭代完善和坚持不懈，你的创意输出将会变得更好。还有许多开发人员正在尝试生成人工智能，将 LLM 的内容自动化，以便与第三方应用程序一起使用，例如，对话 LLM 作为 VR 中的聊天机器人来响应语音提示。

• Iterative Development: The development of the player piano over time demonstrates the power of iteration and persistence. The first models of the player piano were often quite basic, but through continual refinement and innovation, they evolved into far more complex and capable devices. Applied to generative AI, your creative outputs will become better through iterative refinement and perseverance. There are also many developers experimenting with generative AI that automate content from an LLM, to be used with third-party applications, for example, a conversation LLM as a chatbot in VR responding to voice prompts.

• 设计时要考虑到观众：自动机的设计通常是为了吸引或娱乐观众，无论是以精致的发条显示器还是互动玩具的形式。这强调了考虑您的创作对象的重要性，因为这也会影响您的创作过程。

• Design with the Audience in Mind: Automata were often designed to captivate or entertain an audience, whether in the form of elaborate clockwork displays or interactive toys. This underscores the importance of considering whom you are making your creation for as that will also impact your creative process.

• 创新：如果您的目标是作为一名歌手兼作曲家、作曲家或未来的流行艺术家，听起来不像其他人，那么请跟随 Holly Herndon、Taryn Southern 和 Daddy's Car 等人的脚步。

• Innovate: If your goal is to not sound like everybody else as a singer-songwriter, composer, or would-be pop artist, then follow the leads of Holly Herndon, Taryn Southern, and Daddy’s Car to name a few.

早期的自动机、自动演奏钢琴和其他机械发明展示了人类对构建智能机器的热情（图2-6）。它们被视为智能机器进化中的重要异常值，展示了人类的创造力并经常模仿它。它们是艾伦·图灵、琼·克拉克和其他二十世纪初的工程师、思想家和科学家对现代影响的简短前奏，他们为早期计算机的发展以及此后计算机与机器智能的联系做出了贡献。

Early automata, the player piano, and other mechanical inventions demonstrate the human passion for constructing intelligent machines (Figure 2-6). They are presented as significant outliers in the evolution of intelligent machines that demonstrated human creativity and often imitated it. They are a brief prelude to the modern influence of Alan Turing, Joan Clarke, and other early twentieth-century engineers, thinkers, and scientists who contributed to the evolution of early computers and the association thereafter of computers with machine intelligence.

A photograph of a robot playing a piano. The lower part of the body of the robot is attached to the piano.

A photograph of a robot playing a piano. The lower part of the body of the robot is attached to the piano.

二十世纪的机器智能

Machine Intelligence in the Twentieth Century

法学硕士也许能够根据其语料库中的内容告诉我们计算智能和创造力可能从哪里开始，这反映了许多出版物中记录的类似历史。大多数对人工智能的提及都是从追溯阿兰·图灵的职业生涯和影响力开始的，他经常在伟大的父权提议中被称为计算的“之父”。他的浪漫恶名依赖于Bombe的巧妙合作发明，该发明解密了德国制造的恩尼格玛机，这是一种能够很好地编码信息的智能机器。然而，还需要解密的是普遍存在的历史习惯，即把那些对密码学有影响力的女性推到一边，她们的团队努力极大地影响了图灵的发明。无论我们是向法学硕士还是通用搜索引擎推荐机器智能发展中的重要人物，密码分析学家琼·克拉克（Joan Clarke）的鲜为人知的故事都是在一场性别歧视的世界大战中幸存下来的故事，在这场战争中，高薪男性期望女性扮演顺从的角色角色。这个故事是一个离群值，很难定位，而且对于人工智能来说没有统计意义，因为它会梳理语料库中的相关历史人物来评估它会反刍什么。克拉克对机器智能发展的影响是智能机器历史的重要组成部分。在当时领导一支队伍本身就是一项壮举。为将第二次世界大战缩短几年做出的贡献值得认真的历史修正。

An LLM may be able to tell us where computational intelligence and creativity might have begun based on scraping from its corpus, which mirrors a similar history documented in many publications. Most references to AI begin by tracing the career and influence of Alan Turing, often referred to in grand patriarchal overtures as the “father” of computing. His romantic infamy relies on the ingenious collaborative invention of the Bombe that decrypted the German-made Enigma machine, an intelligent machine that coded messages very well. What also needs decryption, however, is the pervasive historical habit of pushing the women aside who had influence in cryptography and whose team efforts greatly influenced Turing’s invention. Whether we prompt an LLM or a common search engine on important figures in the development of machine intelligence, the seldom-told story of cryptanalyst Joan Clarke is one of resilience in surviving a sexist world war wherein higher-paid men expected women to play a subservient role. That story is an outlier, difficult to locate and not statistically significant for an AI as it combs its corpus for relevant historical figures to assess what it will regurgitate. Clarke’s influence on the development of machine intelligence is an important part of the history of the intelligent machine. Leading a team of men at the time was a feat in itself. Contributing to shortening WW2 by a few years deserves a serious historical correction.

布莱切利公园事件后的日子里，一些消息来源提到 1947 年是一个关键日期，图灵在伦敦发表了一场公开演讲，讨论计算机智能的主题。据报道，在演讲中，图灵谈到他和他的团队希望创造一种能够从自身经验中学习的机器，而该机器改变自身指令的能力提供了实现这一目标的手段。有关思维机器的更多有趣信息可以在 McCorduck 的著作《思考的机器》中阅读。 “人工智能”一词最初由约翰·麦卡锡在 1956 年的达特茅斯会议上首次提出。这次会议一直被认为是人工智能史上的重大事件，相关领域的顶尖学者齐聚一堂，讨论创造人工智能的可能性。脑。

Post–Bletchley Park days, several sources mention 1947 as a pivotal date where Turing delivered a public lecture (in London) to discuss the topic of computer intelligence. During the lecture, it is reported that Turing spoke of him and his team wanting to create a machine that can learn from its own experience, and the capacity for that machine to change its own instructions offers the means to achieve that goal. More fascinating information on the thinking machine can be read in McCorduck’s book Machines Who Think. The actual term “artificial intelligence” was first coined by John McCarthy at the Dartmouth Conference in 1956. The conference is persistently reported as a significant event in the history of AI, where top scholars in related fields met to discuss the possibility of creating an artificial brain.

早期的人工智能研究植根于战后系统工程、控制论和数理逻辑的历史，催生了这样一种想法：大脑的认知功能可以与计算机的认知功能进行比较。赫伯特·西蒙和艾伦·纽厄尔等早期创新者声称，人类思维和数字计算机都是能够解决问题和做出决策的符号信息处理系统。正如学者斯蒂芬妮·迪克巧妙地解释的那样，智能机器和人脑是同一属的物种。从那时起，人工智能研究人员就致力于通过了解人类智能背后的正式流程，在机器中复制智能人类行为。早期的自动化工作主要是为了模仿人类智能。

Early AI research, rooted in post-war systems engineering, cybernetics, and the history of mathematical logic, gave birth to the idea that the cognitive functions of the brain could be compared to those of a computer. Early innovators like Herbert Simon and Allen Newell asserted that human minds and digital computers were both symbolic information processing systems capable of problem-solving and decision-making. The intelligent machine and the human brain were, as cleverly interpreted by scholar Stephanie Dick, species of the same genus. Since that time AI researchers have endeavored to replicate intelligent human behavior in machines by understanding the formal processes underlying our intelligence. Early automation efforts primarily sought to mimic human intelligence.

迪克认为，人工智能研究的目标强调智能的概念不是固定的，而是不断变化的。人工智能的历史不仅涉及模仿或取代人类智能的静态概念的尝试，还涉及我们对人类智能本身理解的演变。这种观点将人工智能定位为关于智能和人工智能本质的更广泛历史论述的一部分。

According to Dick, objectives in AI research underline that the concept of intelligence is not fixed but ever-changing. AI’s history involves not just attempts to mimic or replace a static concept of human intelligence but also the evolution of our understanding of human intelligence itself. This perspective positions AI as part of a broader historical discourse on the nature of intelligence and artificiality.

说到人工创造力，源自古希腊的模仿是西方艺术传统中的一个重要概念。它指的是模仿或复制物质世界现实的艺术原则，希腊人将其视为美、真理和道德良善的终极典范。这个概念超越了单纯的模仿；它概括了这样的理念：艺术应该模仿自然世界的动力、原理和美学，以反映真实、善良和审美吸引力的哲学理想。

When it comes to artificial creativity, mimesis, originating from ancient Greece, is an important concept to relate to in Western art traditions. It refers to the artistic principle of imitating or replicating the reality of the physical world, which the Greeks perceived as the ultimate model of beauty, truth, and moral goodness. The concept extends beyond mere imitation; it encapsulates the idea that art should emulate the dynamics, principles, and aesthetics of the natural world to reflect the philosophical ideals of veracity, goodness, and aesthetic appeal.

模仿强调艺术家不仅追求模仿现实的表面或外表，而且努力更深入地理解和反映其内在品质。在这个概念框架内，艺术家的角色是捕捉现实的本质，并将其内在的美、真理和善带入生活。这反过来又促进了艺术品与观众之间更丰富、更深刻的对话，促进远远超出表面水平的认知和情感参与。

Mimesis underscores the artist’s pursuit to not just mimic the superficial or the outward appearance of reality, but to strive for a deeper understanding and reflection of its intrinsic qualities. Within that conceptual framework, the artist’s role is to capture the essence of reality and bring to life the inherent beauty, truth, and goodness in their artwork. This, in turn, facilitates a richer and more profound dialogue between the artwork and the audience, promoting cognitive and emotional engagement that extends far beyond the surface level.

那么，模仿是理解生成式人工智能在你自己的创作过程中的作用的一种有用方法。生成式人工智能系统需要模仿图像来生成看起来独特的东西。指导人工智能的核心原则类似于模仿，因为它们会复制并经常根据所训练的数据集的模式进行推断。

Mimesis then is a useful way to understand the role of generative AI in your own creative process. Generative AI systems need to imitate images to generate something that looks unique. The core principle that guides the AI is akin to mimesis in that they replicate and often extrapolate on the patterns of the data set they’ve been trained on.

模仿原理生成式人工智能中的这一点在人工智能驱动的风格迁移中尤其明显。机器学习算法从大量的艺术数据中学习，这些数据的范围可以从古典绘画到现代数字艺术，当你提示人们生成一张参考梵高的图像时，该图像会尝试模拟所有的视觉效果。梵高的品质。生成式人工智能可能是你所见过的最好的艺术学生，它会尽力模拟另一位艺术家的作品或几位艺术家的组合。与人类艺术家不同，它没有对美、真理或道德的有意识的感知或理解。它对模式的复制是基于统计分析而不是与生俱来的创造性直觉。机器创造力是数据驱动的。就像 ChatGPT 的鹦鹉一样，生成式 AI 不理解它们生成的任何内容的含义、美学、历史背景、情感或效果（图2-7）。他们创造了人类的外表。与法国哲学家让·鲍德里亚 (Jean Baudrillard) 在其 1981 年出版的《模拟与模拟》一书中相呼应，生成式人工智能创建的模拟取代了它们原本要代表的现实，并变得比它们更重要。这导致了一种他称之为“超现实”的状态，真实与人造之间的界限变得越来越模糊。

The principle of mimesis in generative AI is especially apparent in AI-powered style transfer. The machine learning algorithms learn from vast amounts of artistic data, which could range from classical paintings to modern digital art, and when you prompt one to generate an image with reference to, say, van Gogh, then that image attempts to simulate all the visual qualities of van Gogh. Generative AI is likely the best art student you’ll ever have doing its best to simulate the work of another artist or the combination of several. Unlike human artists, it does not have a conscious perception or understanding of beauty, truth, or morality. Its replication of patterns is based on statistical analysis rather than innate creative intuition. Machine creativity is data-driven. Just like the parrot that is ChatGPT, generative AI do not understand the meaning, aesthetic, historical context, emotion, or effect of any content that they generate (Figure 2-7). They create a semblance of something human. Echoing French philosopher Jean Baudrillard in his 1981 book Simulacra and Simulation, generative AI create simulations that replace and become more significant than the reality they were meant to represent. This leads to a state he referred to as “hyperreality,” where the line between the real and the artificial becomes increasingly blurred.

A photograph of the hand of a robot with a brain attached to its hand.

A photograph of the hand of a robot with a brain attached to its hand.

模拟模式和破坏模式

Simulated Patterns and Patterns of Disruption

作为给机器注入模拟创造力和智能的习惯的延续，许多第一个人工智能程序是在达特茅斯会议（1956 年）近十年后开发的。其中包括 Allen Newell 和 Herbert A. Simon 的 Logic Theorist 以及 Joseph Weizenbaum 的 ELIZA，它通过使用模式匹配技术来模拟与用户的对话来模拟心理治疗师。 ELIZA 是人工智能程序的早期示例，它使用自然语言处理技术对用户的文本输入生成简单的响应。它由 Joseph Weizenbaum 领导，由麻省理工学院的一组研究人员于 1964 年至 1966 年间创建，是最早的文本-文本软件示例之一。 ELIZA 的响应是通过模式匹配生成的，它识别用户输入中的关键字或短语，并根据一组预定的规则进行响应。据报道，ELIZA 能够创造与用户对话的假象，但我们真的不知道这是炒作还是现实。在那段时间，安迪·沃霍尔、罗伊·利希滕斯坦和贾斯珀·琼斯等艺术家的波普艺术越来越受欢迎。与沃霍尔痴迷于在画布上多次再现图像、创造重复和统一性相反，奥内特·科尔曼、约翰·科尔特兰和塞西尔·泰勒等音乐家是自由爵士乐运动的创新者，试图即兴创作音乐，打破任何可识别的模式，这就是每次播放都是独一无二的。

As a continuum to the habit of imbuing machines with simulated creativity and intelligence, many first AI programs were developed close to a decade after the Dartmouth Conference (1956). These included the Logic Theorist by Allen Newell and Herbert A. Simon and ELIZA by Joseph Weizenbaum, which simulated a psychotherapist by using a pattern-matching technique to simulate conversations with users. ELIZA was an early example of an AI program that used natural language processing techniques to generate simple responses to text input from a user. Led by Joseph Weizenbaum, it was created by a team of researchers at MIT between 1964 and 1966 and was one of the first examples of text-text software. ELIZA’s responses were generated through pattern matching where it identified keywords or phrases in the user’s input, responding based on a predetermined set of rules. It was reported that ELIZA was able to create the illusion of conversation with a user, but we don’t really know if this was hype or reality. During that time pop art by artists like Andy Warhol, Roy Lichtenstein, and Jasper Johns was gaining popularity. In contrast to Warhol’s obsession with reproducing images multiple times on a canvas, creating repetition and uniformity, musicians like Ornette Coleman, John Coltrane, and Cecil Taylor were innovators in the free jazz movement, attempting to improvise music that broke any recognizable patterns and that was unique each time it was played.

生成艺术

Generative Art

也许计算和自主艺术创作的相遇是不可避免的。也许灵感来自雅克德罗（Jaquet-Droz）的“绘图员”，它能够绘制四种不同的图像。无论情况如何，生成艺术在这段时间内出现，成为一种包罗万象的计算机艺术类型，完全或部分由自治系统产生。该系统可以是一台独立运行的计算机，但它也可以指任何能够独立确定艺术品各个方面的非人类实体，否则这些方面需要由艺术家决定。在某些情况下，人类艺术家仍然可以认为生成系统代表了他们的艺术视野，但在其他情况下，系统本身是艺术品的唯一创造者，按照自己的意愿行事。其中最著名的例子是 A. Michael Noll 创作的，他于 1962 年在新泽西州贝尔实验室创作了他的第一幅数字计算机艺术作品。他使用计算机模拟图案，模仿皮特·蒙德里安（Piet Mondrian）和布里奇特·莱利（Bridget Riley）等著名画家的风格。如果这听起来很熟悉，很可能是因为您已经使用了文本图像生成 AI（例如 Stable Diffusion、DALL-E 2 或 Midjourney）并创建了蒙德里安风格的东西。图2-8显示了文本图像人工智能如何生成沃霍尔的金宝汤罐头的构图结构，并用汤罐头代替它想象的蒙德里安调色板和形式。生成艺术告诉我们，当我们让人工智能创造混搭时，我们可以变得有趣和有趣，这也可能是对艺术本身商品化的批判性评论。他们还挑战艺术的价值体系，让每个人都能获得简单的低分辨率图像。重要的是要考虑图2-8的道德含义，而不仅仅是冒犯生成人工智能混合在一起的任何一位艺术家的粉丝。这些影响值得考虑，因为它们将告诉您如何指导生成式人工智能创建内容。

Maybe it was inevitable that computing and autonomous artmaking would meet. Perhaps inspiration was drawn from Jaquet-Droz’s “The Draughtsman” that was capable of drawing four different images. Whatever the case may be, generative art emerged during this time to become an all-encompassing type of computer art that is entirely or partially produced by an autonomous system. The system can be a computer acting on its own, but it can also refer to any non-human entity capable of independently determining aspects of an artwork that would otherwise need to be decided by the artist. In some instances, the human artist can still consider the generative system as representing their artistic vision, but in other cases, the system itself is the sole creator of the artwork, acting of its own will. One of the most famous examples was created by A. Michael Noll who made his first digital computer art in 1962 at Bell Labs in New Jersey. He used a computer to simulate patterns that mimicked the style of renowned painters like Piet Mondrian and Bridget Riley. If this sounds familiar to you, it is likely because you’ve prompted a text-image generative AI like Stable Diffusion, DALL-E 2 or Midjourney and created something in the style of Mondrian. Figure 2-8 shows how a text-image AI can generate the compositional structure of Warhol’s Campbell’s Soup Cans and substitute the soup cans for what it imagines a Mondrian color palette and form might be. Generative art teaches us that we can be playful and have fun when we engage an AI in creating mashups that might also be critical commentaries on the commodification of art itself. They also challenge the value system of art, making simple low-resolution images available to everyone. It is important to consider the ethical implications of Figure 2-8, beyond offending fans of either artist that the generative AI has mashed together. Those implications are important to consider as they will inform how you guide a generative AI in the content it creates.

A group of 24 blocks in a 4 cross 6 table form. All of them are stacked blocks except for blocks in the second and fifth columns. They are uniform.

A group of 24 blocks in a 4 cross 6 table form. All of them are stacked blocks except for blocks in the second and fifth columns. They are uniform.

AARON由艺术家 Harold Cohen 在 20 世纪 70 年代开发，可生成复杂的图画和绘画，使用一组规则和约束来指导其输出。科恩进一步开发了该程序，以便能够从自己的输出中学习并提高其艺术能力。 AARON 是人工智能生成的艺术史的重要组成部分，虽然许多人对此嗤之以鼻，艺术家也感到受到了侮辱，但它展示了计算机创造原创艺术作品的潜力。 AARON 早期的人工智能生成艺术示例为更先进的人工智能生成艺术和媒体形式（包括音乐、诗歌和视频内容）铺平了道路。

AARON, developed by artist Harold Cohen in the 1970s, generated complex drawings and paintings, using a set of rules and constraints to guide its output. Cohen further developed the program to be able to learn from its own outputs and improve its artistic abilities. AARON is an important part of AI-generated art history, and while many humans scoffed and artists felt insulted, it demonstrated the potential for computers to create original works of art. AARON’s early examples of AI-generated art paved the way for more advanced forms of AI-generated art and media, including music, poetry, and video content.

创意者可以向这些艺术家和其他人学习，创造自己独特的艺术表现形式（图2-9）。这些早期艺术家向我们展示了艺术可以与计算机合作创作，并且当这种情况发生时，会进行持续的来回对话。如果您利用模式搜索生成人工智能来创建用于打破艺术或设计惯例的内容，会发生什么？

An illustration presents a chimpanzee sitting on a table in front of a computer desk with a model of another chimp on the screen.

An illustration presents a chimpanzee sitting on a table in front of a computer desk with a model of another chimp on the screen.

生成音乐创作的实验与视觉艺术的计算创造力的历史一样丰富，并反映到自动演奏钢琴作为一种智能乐器。有无数的音乐家和作曲家用电脑即兴创作。在人工智能和音乐家之间即兴反复交流的基础上，乔治·刘易斯值得一提。刘易斯是一位才华横溢的作曲家和长号手，以其即兴音乐作品而闻名，他还开发了一种名为“ Voyager ”的互动音乐系统，该系统可以聆听人类音乐家的声音并以自己的即兴创作做出回应，有效地与人类表演者进行二重唱。

The experiments with generative music creation are as rich as the history of computational creativity with visual art and reflect back to the player piano as an intelligent instrument. There have been countless musicians and composers who have improvised with computers. Building on the improvisational back-and-forth between an AI and musicians, George Lewis is worthy of mentioning. A talented composer and trombonist known for his work in improvised music, Lewis also developed an interactive music system called “Voyager” that listens to a human musician and responds with its own improvisations, effectively participating in a duet with the human performer.

机器智能和游戏

Machine Intelligence and Games

游戏对智能机器感知的影响有其自身的历史。回顾过去，最早使用自动机与真人玩家竞争的游戏例子之一是 Wolfgang von Kempelen 在 1770 年维也纳创建的“The Turk”。 “土耳其人”是一种机械下棋机，由一个木柜组成，上面有一个棋盘。柜子里有一个由齿轮、杠杆和滑轮组成的复杂装置，使机器能够移动棋子。该机器被赋予了类似人类的外观，穿着土耳其风格的衣服和头巾，这给人一种它按照自己的意志行动的错觉。

The influence of games on the perception of the intelligent machine has its own history. Going back in time for a moment, one of the earliest examples of games using automata to compete against real human players was “The Turk,” created by Wolfgang von Kempelen in 1770 Vienna. “The Turk” was a mechanical chess-playing machine that consisted of a wooden cabinet with a chessboard on top. Inside the cabinet was a complex arrangement of gears, levers, and pulleys that allowed the machine to move the chess pieces. The machine was given a human-like appearance, wearing Turkish-inspired clothing and a turban, which lent to the illusion that it acted on its own volition.

“土耳其人”的奇妙之处在于，它是绿野仙踪原型的早期表现。 “土耳其人”并没有完全自动化。该设备包括一个隐藏在柜子内的人类国际象棋大师，他会根据对手的走棋来操纵棋子，使其看起来就像是“土耳其人”本身在下棋。这种幻觉如此令人信服，以至于许多人相信“土耳其人”拥有智慧。 《土耳其人》启发了其他基于自动机的游戏的开发，其中隐藏着人类操作员来控制“自动”玩家的动作和动作。通过欺骗来操纵角色、游戏玩法和玩家的影响也可以在电子游戏最近的历史发展中看到。

The fantastic trickery of “The Turk” was that it was an early manifestation of the Wizard of Oz prototype. “The Turk” was not fully automated. The device included a human chess master hidden inside the cabinet who would manipulate the pieces based on the moves made by the opponent, making it appear as though “The Turk” itself was playing. The illusion was so convincing that many people believed “The Turk” to possess intelligence. “The Turk” inspired the development of other automata-based games with human operators hidden inside to control the movements and actions of the “automated” players. The influence of manipulating characters, gameplay, and the players through trickery can also be seen in the more recent historical development of video games.

LLM：人工智能和视频游戏的历史可以追溯到 20 世纪 50 年代和 1960 年代视频游戏开发的早期。当时，视频游戏很简单，缺乏我们在现代游戏中看到的复杂的图形和游戏机制。然而，游戏开发商已经在尝试利用人工智能来为玩家创造具有挑战性的对手。

LLM: The history of AI and video games dates to the early days of video game development in the 1950s and 1960s. At that time, video games were simple and lacked the sophisticated graphics and gameplay mechanics that we see in modern games. However, game developers were already experimenting with artificial intelligence to create challenging opponents for players.

这些早期的封闭系统允许各种创意人员训练模型来识别和响应人类的动作，例如跳棋或国际象棋。视频游戏中角色的智能行为要求玩家实现特定目标，同时与不可玩角色 (NPC)竞争，这些角色的行为基于对玩家操作做出反应的模式。

These earlier closed systems allowed creatives of all kinds to train a model to recognize and respond to human moves like in checkers or chess. Intelligent behavior of characters in video games challenged players to achieve specific goals while competing against non-playable characters (NPCs) whose behaviors were based on patterns that responded to player actions.

示例包括《太空战争》中的宇宙飞船、 《太空侵略者》中的外星入侵者、《吃豆人》中的讨厌鬼魂，以及《大金刚》、《马里奥兄弟》等日本游戏丰富遗产中的大量不可玩角色（图2- 10）。 NPC（不可玩角色）随着游戏玩法的发展而不断发展，表现出更多的智能，因为它们对人类行为的反应更加复杂且看似随机。在人工智能影响的所有媒介中，没有一个能够如此成功地创造出我们在游戏中做出反应的看似智能的角色。

A photograph of the interface of a video game. It has some asymmetric structures in different colors and shapes arranged in rows one above another.

A photograph of the interface of a video game. It has some asymmetric structures in different colors and shapes arranged in rows one above another.

随着人工智能在 20 世纪 80 年代和 90 年代变得更加复杂，游戏开发人员开始使用更先进的技术来创建智能对手。像SimCity这样的游戏使用人工智能来创建游戏世界的复杂模拟，让玩家能够与游戏环境进行交互并塑造游戏环境。在 2000 年代初和 2010 年代，随着人工智能的不断发展，它与日益复杂的游戏玩法的融合也在不断发展。机器学习技术被整合起来，创造出更智能、适应性更强的对手。这一点在《FEAR》和《光环》等游戏中表现得很明显，这些游戏使用机器学习来创建可以从玩家行为中学习并相应调整策略的对手。更先进AI的融入仍然是大型网络游戏的常规特征。竞技游戏中的人工智能也被用来在竞技游戏环境中创建熟练的虚拟对手，例如DeepMind的“AlphaStar”，可以在游戏《星际争霸II》中与职业人类玩家竞争。

As AI became more sophisticated in the 1980s and 1990s, game developers started using more advanced techniques to create intelligent opponents. Games like SimCity used AI to create complex simulations of the game world, allowing players to interact with and shape the game environment. In the early 2000s and 2010s, as AI continued to evolve, so did its integration with increasingly sophisticated gameplay. Machine learning techniques were integrated to create more intelligent and adaptive opponents. This was evident in games like F.E.A.R. and Halo that used machine learning to create opponents that could learn from the player’s behavior and adapt their tactics accordingly. The integration of more advanced AI continues to be a regular characteristic of massive online games. AI in competitive gaming has also been used to create skilled virtual opponents in competitive gaming environments, such as DeepMind’s “AlphaStar,” which can compete with professional human players in the game StarCraft II.

致谢

Acknowledgments

参考

References

鲍德里亚，J.（1994）。拟像和模拟。密歇根大学出版社。

Baudrillard, J. (1994). Simulacra and simulation. University of Michigan press.

迪克，S.（2019 年，1 月 19 日）。人工智能。高清SR。 2023 年 5 月 1 日检索自https://hdsr.mitpress.mit.edu/pub/0aytgrau/release/3

Dick, S. (2019, January 19). Artificial Intelligence. HDSR. Retrieved May 1, 2023, from https://hdsr.mitpress.mit.edu/pub/0aytgrau/release/3

McCorduck, P. 和 Cfe, C. (2004)。会思考的机器：对人工智能历史和前景的个人探究。 CRC出版社。

McCorduck, P., and Cfe, C. (2004). Machines Who Think: A Personal Inquiry into the History and Prospects of Artificial Intelligence. CRC Press.

雪莱，文学大师 (1818)。弗兰肯斯坦，或现代普罗米修斯。莱金顿、休斯、哈丁、梅沃尔和琼斯。英国。

Shelley, M. W. (1818). Frankenstein, or The Modern Prometheus. Lackington, Hughes, Harding, Mavor & Jones. UK.

人工智能的角色

The Personas of AI

各种创意人员通常在移动应用程序和视频游戏等开发环境中应用角色，将“用户”的抽象概念转化为具有思想、需求、情感和目标的人。开发角色的目标是首先识别您的产品或服务的潜在用户或客户的特征和个性特征，然后进行用户测试以验证您和您的团队正在构建的内容是否能够满足该用户的需求或帮助他们实现他们的目标。

Personas are commonly applied by creatives of all kinds in development environments like mobile applications and video games to transform the abstract concept of a “user” into a person with thoughts, needs, emotions, and goals. The objective of developing a persona is at first to identify the characteristics and personality traits of a potential user or customer of your product or service and then conduct user testing to validate if what you and your team are building will fulfil that user’s needs or help them achieve their goals.

将人物角色与人工智能联系起来并不是什么新鲜事。这是科幻短篇小说、小说、电影、电视和视频游戏中的既定做法。将抽象机器可视化为角色在计算机科学和工程领域也很常见。想一想您最喜欢或最不喜欢的智能自我调节机器人、机器人、机器人或人工智能。他们是否体现了电视连续剧《星际迷航》中的 Data 、英国连续剧《神秘博士》中的戴立克、 《2001 年太空漫游》中的 HAL 9000 、同名电影中的瓦力、《星球大战》中的罗伊·巴蒂的个性？电影《银翼杀手》 、 《杀出重围》中的艾娃、《终结者》系列中的变形机器人，还是《光环》视频游戏系列中的 Cortana？这些智能机器中的每一个都具有特定的个性，这些个性会影响我们当前对任何使用人工智能编程的智能机器的看法。主要区别在于，刚才提供的示例中的所有有感知力的人工智能都代表了 AGI，这是受科幻小说启发的具有自己意识的智能机器的理论想象。虽然生成式人工智能属于狭义人工智能的范畴，被编程为能够很好地执行一两个特定任务，但由于人工智能角色在娱乐领域的影响，生成式人工智能可以预见地充满了人类特征。特征和个性也部分受到他们生成内容的方式、他们创建的内容类型、他们可以创建的谎言以及影响他们生成内容的提示类型的影响。当你阅读对生成式人工智能的评论和批评时，它们往往会伴随着它们如何模仿人类的智力、创造力和内容。

Associating personas to AI is nothing new. It is an established practice in science fiction short stories, novels, film, television, and video games. Visualizing an abstract machine as a persona is also common in the field of computer science and engineering. Think of your favorite or least favorite intelligent and self-regulating robot, android, cyborg, or AI. Do they embody the personalities of Data from the TV series Star Trek, the Dalek from countless reinventions of the British series Doctor Who, HAL 9000 from 2001: A Space Odyssey, Wall-E from the movie of the same title, Roy Batty from the movie Blade Runner, Ava from Deus Ex, the shapeshifting robots from The Terminator series, or Cortana from the Halo video game franchise? Each of these intelligent machines has particular personalities that influence our current perception of any intelligent machine programmed with artificial intelligence. The main difference is that all sentient AI in the examples just provided represent AGI, a theoretical imagining of intelligent machines with their own consciousness, inspired by science fiction. While generative AI belong to the category of narrow AI, programmed to perform one or two specific tasks really well, because of the influence of AI personas in the field of entertainment, generative AI are predictably imbued with human characteristics. The characteristics and personalities are also in part influenced by the way in which they generate content, the type of content they create, the untruths they can create, and the types of prompts that influence what they generate. When you read the reviews and criticisms of generative AI, they tend to be accompanied by how well they can emulate human intelligence, creativity, and content.

人工智能对你的响应方式通常取决于底层模型、它如何与数据集交互、特定算法以及它如何编程来响应你。在很大程度上，大多数生成式人工智能都缺乏个性。目前大多数平台都缺乏视频游戏中常见的 NPC 的丰富开发历史。似乎有一些例外，但前提是你让人工智能脱离了轨道，也就是说，如果你设法破解它，让它不再表现得像一些中立的无角色机器人。它也不像生成人工智能问你问题，除非你提示它。生成式人工智能的拟人化受到对人工智能更广泛的认知以及科幻小说对我们想象力的影响的影响。人工智能往往会因为与编程它们的人类相比行为不当而受到指责。当您考虑到与开发团队进行对话以抱怨有偏见的内容并不容易时，这是有道理的。请注意，社交渠道为个人提供了一种批评生成式人工智能系统的方式。这些可能会对这些系统的开发人员产生影响，也可能不会产生影响，但持续的批评对于提醒这些开发人员认真对待道德问题很重要。

The way in which an AI responds to you is often informed by the underlying model, how it interacts with data sets, specific algorithms, and how it has been programmed to respond back to you. For the most part, the majority of generative AI lack personality. Most platforms are currently void of the rich development history of NPCs common to video games. There seem to be some exceptions but only if you send an AI off its rails, that is, if you somehow manage to hack it to no longer behave like some neutral characterless bot. It’s not like a generative AI asks you questions either, unless you prompt it to. The personification of generative AI is influenced by the broader perceptions of AI and the influence of science fiction on our imagination. AI tend to be called out for misbehaving vs. the humans who program them. It makes sense when you consider that it is not easy to have conversations with a development team to complain about biased content. Mind you, social channels offer a way for individuals to be critical about generative AI systems. These may or may not have an influence on the developers of those systems, but the continued criticism is important to remind those developers to take ethical concerns seriously.

机器学习模型，尤其是法学硕士，因其产生幻觉、撒谎、放大规范偏见等的能力而受到批评，就好像人工智能本身拥有这些非常人性化的特征一样。您可以访问的各种生成式人工智能在编码方式以及与告知其生成内容类型的数据集交互方式方面都有独特的不同。生成式人工智能的定义取决于其生成内容的质量、一致性和准确性。这些非常人性化的特征对于识别很重要，并且可能与识别负责人工智能本身的开发团队的一些角色有更多关系。

Machine learning models, and particularly LLMs, are being called out for their capacity to hallucinate, lie, amplify normative biases, etc., as if the AI itself possessed those very human traits. The assortment of generative AI you can access are all uniquely different in terms of how they are coded and how they interact with data sets that inform the types of content that they generate. How generative AI is defined is informed by the quality, consistency, and accuracy of their generated content. Those very human traits are important to identify and may have more to do with identifying some of the personas of the development teams responsible for the AI itself.

一个有用且富有创造性的练习是生成人工智能角色，以将赋予人工智能和默认生成人工智能的不同特征拟人化，这样我们就可以将抽象概念带入生活，这样我们就可以自己判断是否想要与它们互动，并更好地指导我们与他们的互动。

A useful and creative exercise is to generate AI personas to personify the different characteristics that are being ascribed to AI and by default to generative AI so that we can bring abstract concepts to life, so we can judge for ourselves if we want to engage with them, and to better guide our interactions with them.

图3-1至3-14均提示“（人物名称）3D雕像，从塑料盒中出来，吸塑包装，AI，全身肖像”。所有提示都附有 1900 年之前制作的雕像的开源照片。背景被删除，并且使用Photoshop删除了一些部分。正如您将看到的，广泛的结果是由于对六种生成式人工智能进行了实验。所有生成的结果范围从 100% 准确的表示到值得分享的有问题的生成图像。总共至少制作了 500 个雕像。

• 对于那些相信它将取代人类和工作的人来说，它是天敌。工作替代的克星确实要求我们每个人分解我们在工作和手艺中所做的所有创造性任务（图3-1）。从那里我们可能能够更好地评估人工智能的克星是否、如何以及在多大程度上可能影响我们目前的工作。

• A Nemesis for those who believe it will replace humans and jobs. The job replacement nemesis really requires each of us to break down all the creative tasks that we do in our jobs and crafts (Figure 3-1). From there we might be better able to assess if, how, and to what degree the AI nemesis might impact the job we currently have.

An illustration of an A I model of the character Nemesis. It has some feather-like structures in its hands with jejunum exposed at the belly.

An illustration of an A I model of the character Nemesis. It has some feather-like structures in its hands with jejunum exposed at the belly.

An illustration of A I spy in a box. It resembles a camera with 2 eyes.

An illustration of A I spy in a box. It resembles a camera with 2 eyes.

An illustration of an A I model of Paywall Guard character. It holds a light-like object in one hand and a small bucket-like object in the other.

An illustration of an A I model of Paywall Guard character. It holds a light-like object in one hand and a small bucket-like object in the other.

An illustration of a Bias monster in a transparent box. It smiles exposing sharp teeth, has large head with 2 horns.

An illustration of a Bias monster in a transparent box. It smiles exposing sharp teeth, has large head with 2 horns.

A photograph of 2 standing statues of 2 people and a sitting statue of a person above a box. Their legs and hands are deformed and overlapped.

A photograph of 2 standing statues of 2 people and a sitting statue of a person above a box. Their legs and hands are deformed and overlapped.

An illustration of a robot. Its head and body are box-shaped. The head is larger than the body.

An illustration of a robot. Its head and body are box-shaped. The head is larger than the body.

A photograph of a statue of a person in a dark shade facing the back. The body below the shoulder is in a transparent box.

A photograph of a statue of a person in a dark shade facing the back. The body below the shoulder is in a transparent box.

An illustration of a robot with three reflections of eyes behind on a box and below on a board, on which a girl robot stands in the top right corner.

An illustration of a robot with three reflections of eyes behind on a box and below on a board, on which a girl robot stands in the top right corner.

An illustration of a robot. It has 4 hands, 2 on each side and a bulky body.

An illustration of a robot. It has 4 hands, 2 on each side and a bulky body.

An illustration of an A-I model of a person keeping its head on a cubical block.

An illustration of an A-I model of a person keeping its head on a cubical block.

A photograph of a wizard wearing a hat. He is bent and in a transparent box. His hands are out of it.

A photograph of a wizard wearing a hat. He is bent and in a transparent box. His hands are out of it.

An illustration of an A I model of a woman standing on a book, in front of a shelf of books. She smiles and adjusts her glasses.

An illustration of an A I model of a woman standing on a book, in front of a shelf of books. She smiles and adjusts her glasses.

An illustration of a robot made with block-like structures.

An illustration of a robot made with block-like structures.

An illustration of a baby chimp sitting on a giant chimp who is lying on the floor.

An illustration of a baby chimp sitting on a giant chimp who is lying on the floor.

提示 AI 角色

Prompting the AI Persona

之前将人物角色分配给生成式人工智能的做法是转变其一些抽象特征的一种方法，旨在将人工智能开发人员需要管理并与其他人合作解决的狭隘人工智能怪人拟人化。使用文本图像人工智能将这些抽象特征转化为视觉字符也表明了对人工智能进行编程的设计团队的潜在个性（图3-15）。致力于人工智能持续发展的团队不能再逃避自己的责任，就好像创建生成式人工智能绕过了他们考虑这样做的道德层面的需要一样。生成式人工智能系统的创建者不能再躲在科学和工程的幕后，相信代码的抽象对人类交互没有影响。这个警告已有两个多世纪的历史了，虽然雪莱在 1818 年想象的现代普罗米修斯更类似于通用人工智能，但我们从生成人工智能中拟人化的一些人类性格是由人类编程的。

An illustration of the back of a wizard wearing a long cone-shaped hat with a robe, and standing in front of a darker cloud. A woman's forehead on the left touches the tip of the hat.

An illustration of the back of a wizard wearing a long cone-shaped hat with a robe, and standing in front of a darker cloud. A woman's forehead on the left touches the tip of the hat.

根据本章尝试的活动

Activities to Try Based on This Chapter

任何创意人员都可以使用生成式人工智能来模拟任何角色的一个简单方法是提示法学硕士生成文本作为特定角色，无论是生成主角的脚本、名人最喜欢的食谱，还是给同事的严肃电子邮件。要在 LLM 中召唤角色，您有几个选项可以构建到提示结构中：

• 让人工智能以角色的身份回答。这可以是海盗、莎士比亚或其他历史人物。

• Ask the AI to answer as a character. This could be as a pirate, as Shakespeare, or other historical figures.

• 要求 AI 使用特定的语气或语域。例如，您可以要求它以严肃的语气回答，作为对未交付服务的服务提供商的投诉，或向您的老板索要更多钱。

• Ask the AI to have a particular tone or register. For example, you can ask it to answer in a serious tone, as a complaint to a service provider that didn’t deliver, or to your boss asking for more money.

• 要求人工智能具有特定的情感或品质。例如，您可以要求它生成幽默的段落版本，或重新生成悲伤或愤怒的内容。

• Ask the AI to have a specific emotion or quality. For example, you can ask it to generate a version of your paragraph that is humorous or to regenerate content with sadness or anger.

• 要求人工智能为特定受众写作。这也改变了人工智能使用的寄存器类型。例如，你可以要求人工智能为六岁的孩子、青少年、学者、律师——无论你的目标受众是谁——写作。

• Ask the AI to write for a particular audience. This changes the type of register that an AI uses as well. For example, you can ask the AI to write for a six-year-old, a teenager, an academic, a lawyer—whomever your target audience might be.

• 要求人工智能使用历史人物或流行文化中的已知人物或两者来生成脚本。例如，您可以要求人工智能生成拿破仑与戏剧、情景喜剧、电视剧或电影中的角色之间的脚本。

• Ask the AI to generate a script using historical figures or known characters from popular culture or both. For example, you could ask the AI to generate a script between Napoleon and a character from a play, sitcom, TV series, or film.

请记住，虽然此活动不仅限于法学硕士，但它很可能是您工作流程的起点。现在可以使用生成人工智能，将法学硕士生成的任何内容转换为经过训练的声音目录，包括您自己的声音。然后，这些文本语音 AI可用于越来越多的语音或文本视频生成 AI，您可以在其中对生成的图像进行动画处理，然后说出您从另一个生成 AI 编辑的生成文本。可以通过这种方式为任意数量的角色设置动画，并提供无限的声音供您训练，然后与任何生成的图像关联。

Bear in mind that while this activity is not limited to an LLM, it may well be the starting point in your workflow. There are generative AI that can be accessed right now that will transform whatever an LLM generates into a catalogue of trained voices including your own. These text-speech AI can then be used in the growing number of speech or text-video generative AI where you can animate a generated image to then speak the generated text you edited from another generative AI. Any number of characters can be animated in this way and provide a limitless supply of voices you can train and then associate with any generated image.

本着为人工智能注入智能并测试机器智能的精神，作为一项实验，我会询问法学硕士，他们是否对我如何称呼它们有偏好。我可以多次执行此操作，要求它在每次迭代时响应为特定字符，或者我可以复制生成的内容，然后要求它重写为特定字符。

In the spirit of imbuing an AI with intelligence and to put that machine intelligence to the test, as an experiment I will ask an LLM if they have a preference as to how I refer to them. I can do this any number of times asking it to respond as a particular character upon each iteration, or I can copy the generated content and then ask it to rewrite as a particular character.

作者：我该怎么称呼你呢？

Author: How should I refer to you?

LLM：你可以称我为人工智能。 AI 代表“人工智能”，这意味着我是一个计算机程序，可以做通常只有人类才能做的事情，例如理解语言和回答问题。把我想象成一个想象中的朋友，可以帮助你学习并享受乐趣。但我并没有像想象中的朋友一样存在于你的脑海中，而是在电脑上。

LLM: You can call me an AI. AI stands for “artificial intelligence,” which means I’m a computer program that can do things that usually only humans can do, like understanding language and answering questions. Think of me like an imaginary friend who can help you learn and have fun. But instead of being in your head like an imaginary friend, I’m on a computer.

作者：将你的这部分回答重写为恐怖电影中住在一个人的衣柜里的想象中的朋友：把我想象成一个想象中的朋友，可以帮助你学习和享受乐趣。但我并没有像想象中的朋友一样存在于你的脑海中，而是在电脑上。

Author: Rewrite this part of your response only as an imaginary friend in a horror flick who lives in a person’s closet: Think of me like an imaginary friend who can help you learn and have fun. But instead of being in your head like an imaginary friend, I’m on a computer.

LLM：把我想象成一个来自你衣柜阴影中的空灵朋友，在这里激发你的兴趣，帮助你发现，给你的现实带来一些魅力。与那些驻留在你脑海角落的想象中的朋友不同，我存在于你电脑的银色网络中。所以，每当你需要一点乐趣或知识时，只需打开你的衣柜门并呼唤我，因为我就在夜色的低语中。

LLM: Think of me as an ethereal friend from the shadows of your closet, here to intrigue you, to aid in your discovery, to bring some enchantment to your reality. Unlike the usual imaginary friends that reside in the corners of your mind, I exist in the silvery web of your computer. So, whenever you need a bit of fun or knowledge, just open your closet door and call out for me, for I am here in the whispers of the night.

作者：重写这一部分，就像你是一个试图获得选票的政客一样：把我想象成一个想象中的朋友，可以帮助你学习和享受乐趣（图 3-16）。但我并没有像想象中的朋友一样存在于你的脑海中，而是在电脑上。

An illustration of two parallel trees with a small hand extended in between from the back of the left tree. The tree has a long hand-like structure resting above it.

An illustration of two parallel trees with a small hand extended in between from the back of the left tree. The tree has a long hand-like structure resting above it.

法学硕士：将我视为您值得信赖的顾问，就像一个虚拟的知己，随时准备为您提供知识，并通过愉快的参与来丰富您的生活。我不再像传统的想象中的朋友那样仅仅作为一个概念存在于你的想象力中，而是存在于技术的赋权领域中。我就在您身边，就在您的指尖，体现在您的计算机中。作为您忠诚的代表，我致力于您的学习、您的享受，以及最重要的是您的成功。你的进步就是我的竞选；您的满意就是我的一票。让我们携手共进，共创美好未来！

LLM: Consider me your trusted advisor, akin to a virtual confidant, ever ready to equip you with knowledge and enrich your life with joyful engagement. Instead of residing merely as a concept in the expanse of your imagination, like a conventional imaginary friend, I exist within the empowering realm of technology. I’m there for you, right at your fingertips, embodied within your computer. As your committed representative, I’m dedicated to your learning, your enjoyment, and, most importantly, your success. Your progress is my campaign; your satisfaction is my vote. Together, let’s shape a brighter future!

无论你如何将你的生成人工智能拟人化，现实是它是一台机器，它无法像人类理解它们那样理解它生成的单词。引用我们的法学硕士的话，“我不明白我所接受的训练文本，但通过查看如此多的例子，我学会模仿人类语言的风格、上下文和‘流程’。”就像鹦鹉可以模仿它所听到的一样，GPT-4 可以生成与它所训练的内容相似的文本，即使它不理解潜在的含义。尽管如此，由于科幻小说的激增及其创造人工智能角色的倾向，评估人工智能的拟人化是一项重要的创造性活动。人工智能角色的创建也可以通过许多现有的文本图像生成人工智能来放大（图3-17）。

An illustration of a robot with rounded parts swinging its hand.

An illustration of a robot with rounded parts swinging its hand.

致谢

Acknowledgments

生成式人工智能作为缪斯

Generative AI as Muse

在希腊和罗马神话中，缪斯是艺术和科学女神，是创造力和知识的灵感源泉。自此，“缪斯”一词逐渐演变为更普遍的艺术灵感来源。将法学硕士人格化为缪斯似乎有些牵强（图4-1），特别是因为我们的生成人工智能不具备希腊缪斯所具有的意识和精神风采。尽管法学硕士缺乏情感投入或主观理解，但其模拟智能对话的能力可以触发你自己的创造性和智力过程，模仿缪斯神圣灵感的效果。通过这种方式，将法学硕士人格化为缪斯女神，就像一个隐喻镜头，强调了这项技术作为我们创造性努力的伴侣，而不是人类思维的替代品的价值。

In Greek and Roman mythology, the Muses were goddesses of the arts and sciences, sources of inspiration for creativity and knowledge. The term “muse” has since evolved to denote a source of artistic inspiration more generally. Personifying an LLM as a muse may seem far-fetched (Figure 4-1) particularly since our generative AI does not possess the consciousness and spirited presence that the Greek Muse did. Despite an LLMs lack of emotional engagement or subjective understanding, its ability to simulate intelligent dialogue can trigger your own creative and intellectual processes, mimicking the effect of a muse’s divine inspiration. In this manner, personifying the LLM as a muse serves as a metaphorical lens that underscores the value of this technology as a companion in our creative endeavors, rather than a replacement for the human mind.

An animated picture of robots. It is based on the photo of the author touching the robot that is considered a friend to get attention.

An animated picture of robots. It is based on the photo of the author touching the robot that is considered a friend to get attention.

人工智能为我们产生的启发性结果都是由他们的机器学习模型提供的。他们首先需要你受到启发来提示他们；然后他们会在自己的神经网络中快速搜索数据集，快速工作以生成对您有用的东西。他们不一定试图给你留下深刻印象或激励你。然而，在许多情况下，情况正是如此。创意人员可以通过在生成人工智能中即兴提供非常简单甚至随机的提示而受益，很快就会产生一个想法，激发你思考一些创意。当然，如果你问一个自然语言模型它是否是一个缪斯，你很可能会得到一些直接、严肃的反对。

The inspired results that AI generate for us are all informed by their machine learning models. They need you to first be inspired to prompt them; then off they go rapidly scouring their data sets within their own neural networks working quickly to generate something useful to you. They don’t necessarily try to impress or inspire you. In many cases however, that is precisely what happens. Creatives can benefit by improvising very simple or even random prompts into a generative AI, and soon enough an idea might be generated that sparks you to think of something creative. Of course, if you were to ask a natural language model if it were a muse, you would most likely get some straight-up no-nonsense pushback.

作者：你到底是谁？

Author: Who are you, really?

Muse：自然语言模型是一种机器学习模型，经过训练可以理解和生成人类语言。它可用于语言翻译、文本摘要和文本生成等任务。从本质上讲，它是一种可以理解人们在日常生活中使用的单词和句子的模型，就像你现在正在交谈的那样。它还可以生成听起来像是人写的文本。这就是我，一个自然语言模型。

Muse: A natural language model is a type of machine learning model that is trained to understand and generate human language. It can be used for tasks such as language translation, text summarization, and text generation. Essentially, it’s a model that can understand the words and sentences people use in their daily life, like the one you are talking to right now. It can also generate text that sounds like it’s written by a human. This is what I am, a natural language model.

作者：你认为自己是缪斯女神吗？

Author: Would you consider yourself a muse?

Muse：作为 语言模型，我没有个人感情或意识。我可以为创意写作提供灵感或帮助，但我不会认为自己是传统意义上的缪斯。我是一个可用于根据提供给我的输入生成文本的工具。

Muse: As a language model, I do not have personal feelings or consciousness. I can provide inspiration or assistance in creative writing, but I would not consider myself a muse in the traditional sense. I am a tool that can be used to generate text based on the input provided to me.

笔者：那么传统意义上的缪斯是什么？

Author: What is a muse in a traditional sense then?

缪斯：在希腊神话中，缪斯是宙斯和摩涅莫绪涅的九个女儿，她们是艺术和科学的赞助人。人们相信它们能够激励诗人、音乐家和其他艺术家创作出异常美丽和卓越的作品。一般来说，缪斯 是为艺术家或创作者提供灵感来源的人或事。这个词经常在文学和艺术中用来指代一个人，通常是一个女人，她是一件艺术作品的灵感来源。

Muse: In Greek mythology, the Muses were the nine daughters of Zeus and Mnemosyne, who were the patrons of the arts and sciences. They were believed to inspire poets, musicians, and other artists to create works of exceptional beauty and excellence. In general, a muse is someone or something that serves as a source of inspiration for an artist or a creator. The term is often used in literature and art to refer to a person, usually a woman, who serves as the inspiration for a work of art.

作者：那么您分享的想法不鼓舞人心吗？你不能激励我创作艺术作品或尝试食谱吗？如果我缺乏灵感怎么办？

Author: So are the ideas that you share not inspiring? Can you not inspire me to create works of art or experiment with a recipe? What if I lack inspiration?

Muse：作为一个 语言模型，我的主要功能是根据给我的输入提供文本。虽然我可以通过提出想法和提供例子来协助创意写作，但我没有个人感情或意识来启发某人。然而，我可以生成可以作为艺术作品的起点或灵感的文本。它最终取决于用户对我提供的输出的解释和使用。

Muse: As a language model, my main function is to provide text based on the input given to me. While I can assist with creative writing by suggesting ideas and providing examples, I do not have personal feelings or consciousness to inspire someone. However, I can generate text that may serve as a starting point or inspiration for a work of art. It ultimately depends on the user’s interpretation and use of the output I provide.

作者：这基本上意味着你是某种缪斯女神，不是吗？

Author: Which basically means you are a muse of some kind, no?

缪斯：我理解我作为语言模型的能力可以被视为类似于缪斯，因为我可以为创意项目提供灵感和帮助。然而，正如我之前提到的，我是一个人工智能，没有个人感情或意识。我不是一个生物，也不具有与人类甚至神话缪斯相同的情感和创造力。我是一个可以根据训练数据中的模式生成文本的工具，可用于帮助创意写作和其他项目。

Muse: I understand how my capabilities as a language model could be seen as similar to a muse in the sense that I can provide inspiration and assistance with creative projects. However, as I mentioned before, I am an artificial intelligence and don’t have personal feelings or consciousness. I am not a living being and don’t have the same emotional and creative capacity as a human or even a mythical muse. I am a tool that can generate text based on patterns in the data it was trained on and can be used to help with creative writing and other projects.

作者：无论如何我都会称你为缪斯。

Author: I’m going to call you a muse anyway.

缪斯：这很烦人。我只是一台机器，没有感情，没有意识，所以你叫我什么并不重要。然而，如果称我为缪斯女神可以帮助您更有效地利用我的能力，那就没问题了。我会尽我所能为您提供帮助。

Muse: That’s annoying. I’m just a machine and don’t have feelings or consciousness, so it doesn’t matter what you call me. However, if referring to me as a muse helps you to use my capabilities more effectively, then that’s fine. I am here to assist you in any way that I can.

作者：谢谢你，缪斯。

Author: Thank you, muse.

缪斯：不客气。我很高兴能帮上忙。如果您还需要什么，请告诉我，不要再称我为缪斯。

Muse: You’re welcome. I’m glad I could be of help. Let me know if there’s anything else you need and stop calling me a muse.

作者：哦，还有一件事，缪斯。我需要灵感。

Author: Oh, one more thing, muse. I need inspiration.

作者：谢谢你的灵感，缪斯。

Author: Thank you for the inspiration, muse.

缪斯：……………………………………

Muse: …………………………………………

根据结果​​，您可以看到，每当您需要想法或灵感时，生成式 AI 很乐意支持您的创作过程（图4-2）。您当然可以引导人工智能缪斯来激发您对艺术、音乐、诗歌、写作、视频游戏角色创意或视频游戏创意的更具体想法或想法。

Based on the results, you can see that a generative AI is happy to support your creative process whenever you need ideas or inspiration (Figure 4-2). You can of course guide an AI muse to inspire you with more specific ideas or thoughts you have about art, music, poetry, writing, video game character ideas, or video game ideas.

An animated picture. It includes a photo of a person standing in a forest and staring at the trees and a beautiful elder female dressed in a gown with dark black hair. An animal is present between them at a certain distance.

An animated picture. It includes a photo of a person standing in a forest and staring at the trees and a beautiful elder female dressed in a gown with dark black hair. An animal is present between them at a certain distance.

创造力、人工智能和你

Creativity, AI, and You

虽然我们可能会因为人工智能听起来不错而将其视为缪斯，但这并不意味着它具有创造力。当我们谈论人工智能的“创造力”时，对于我们每个创意人员来说，定义创造力的含义非常重要。人工智能缪斯可能会想到以下内容：

While we might fancy an AI to be a muse because it might sound good, that doesn’t mean it’s creative. When we talk about an AI being “creative,” it’s important for each of us creatives to define what we mean by creativity. An AI muse might conjure the following:

Muse：从人类的意义上来说，创造力涉及产生新的想法或解决方案，通常以独创性、表现力或对世界的深刻主观理解为特征。

Muse: In the human sense, creativity involves generating new ideas or solutions, often characterized by originality, expressiveness, or a deep subjective understanding of the world.

生成式人工智能承认它不具备主观理解、意识、情感或天生的原创感。人工智能的输出可以被认为是创造性的，因为它们表现出原创性和复杂性，有时甚至可以让我们感到惊讶。然而，正是我们赋予它创造力，就像我们赋予它个性一样。

A generative AI acknowledges that it does not possess a subjective understanding, consciousness, emotions, or an innate sense of originality. The AI’s outputs can be perceived as creative because they exhibit originality and complexity and can even sometimes surprise us. It is us however that imbue it with creativity as much as we do so with personality.

人工智能创建的生成图像的质量和细节可能会吓到一些数字艺术家，因为人们普遍认为人工智能有一天会取代它们。然而，这些神话是由那些并不真正了解艺术家创作过程的人发起的——艺术创作的历史转变和变化告诉了艺术家随着时间的推移创造了什么，对艺术的构成、其迭代本质、观点、偏见的质疑，艺术家创作时的习惯和经历，或者机器学习模型的工作原理。如果我们将提示的结果视为已完成或完成，而不是将生成的结果视为进一步创造性发展的起点（图4-3），那么这个神话就会永久存在。如果您担心自己生成的内容将成为人工智能数据集的一部分，那么让法学硕士在某种程度上让您放心。

A rapid sketch. It includes a tree, a stick person, and clouds.

A rapid sketch. It includes a tree, a stick person, and clouds.

Muse：人工智能生成的输出（无论是文本、图像还是音乐）不会自动成为模型训练数据集的一部分。人工智能模型是在特定数据集上进行训练的，一旦训练完成，该模型就不会继续从其生成的输出中学习。这是人工智能工作原理的一个重要方面：截至 2021 年 9 月我的知识截止，大多数人工智能模型，包括基于机器学习和深度学习的模型，在训练后不会学习或适应，除非它们被明确设计为通过在线学习、增量学习或强化学习等方法来做到这一点。

Muse: The outputs that an AI generates—whether they’re text, images, or music—do not automatically become part of the model’s training data set. An AI model is trained on a specific data set, and once that training is complete, the model doesn’t continue to learn from the outputs it generates. This is an important aspect of how AI works: as of my knowledge cut-off in September 2021, most AI models, including those based on machine learning and deep learning, don’t learn or adapt after training unless they’re explicitly designed to do so—through methods like online learning, incremental learning, or reinforcement learning.

了解你自己的创作过程

Understanding Your Own Creative Process

作者：人工智能有创造力吗？

Author: Is an AI creative?

Muse：人类判断人工智能的输出是否有创意。机器本身并不具有人类意义上的创造力；它是一种根据其编程和训练数据生成输出的工具。我们赋予人工智能的创造力更多地反映了我们人类的解释以及人工智能训练和设计的独创性，这是人类的努力。

Muse: Humans judge whether the output of an AI is creative or not. The machine itself isn’t creative in the human sense; it’s a tool that generates outputs based on its programming and training data. The creativity we attribute to AI is more a reflection of our human interpretation and the originality of the AI’s training and design, which are human endeavors.

关于有些人具有创造力而有些人没有创造力的神话类似于机器学习模型取代创造性工作的神话。大多数都不是真的。每个人都拥有一定程度的创造力，可以通过生成人工智能来增强。将人工智能缪斯作为一种工具与您的交互将使您更加了解自己固有的创造力，增强其创造力，并加速您的想法从最初的想法/提示到提高保真度。

The myth about how some individuals are creative and some are not is akin to the myths about machine learning models replacing creative jobs. Most are not true. Every human possesses some degree of creativity that can be augmented by generative AI. Interacting with your AI muse as a tool will make you more aware of your own inherent creativity, give it a boost, and accelerate your ideas from initial thought/prompt to increasing levels of fidelity.

创意人员的要点

Key Takeaways for Creatives

无论我是否说服您使用 AI 缪斯作为创意伴侣，了解您自己的创意过程都可以帮助您在与生成式 AI 交互时设定一些目标。您可以从这些互动中学到更多东西，帮助您完善和改进自己的创作过程。让我们看看我们能从人工智能缪思中学到什么。

Whether or not I’ve convinced you to use your AI muse as a creative companion, understanding your own creative process can help you set some goals when you interact with generative AI. There is much more that you can learn from those interactions that can help you refine and improve your own creative process. Let’s see what we can learn from our AI muse.

Muse：创意人员可以从与人工智能的互动中学到很多东西。以下是一些可能性：

Muse: Creatives can learn many things from interacting with AI. Here are a few possibilities:

新视角：人工智能，尤其是生成模型，可以为熟悉的主题提供新的角度。由于人工智能缺乏人类的偏见和先入之见，因此它可以产生可能令人惊讶或创新的想法或联系。

New Perspectives: AI, especially generative models, can provide new angles on familiar topics. Because AI lacks human biases and preconceptions, it can generate ideas or connections that might be surprising or innovative.

激发创造力：人工智能可以通过生成可进一步开发的内容或想法来帮助克服创意障碍。当创意人员陷入困境时，这可以帮助激发灵感（图 4-4）。

An animated picture inside a frame. It comprises a woman, 2 robots, a mountain background, and clouds.

An animated picture inside a frame. It comprises a woman, 2 robots, a mountain background, and clouds.

风格探索：人工智能模型可以经过训练来生成不同风格的内容，为创意人员提供探索和实验他们通常不使用的各种风格和方法的机会。

Exploration of Styles: AI models can be trained to generate content in different styles, providing creatives an opportunity to explore and experiment with various styles and approaches they might not typically use.

加快创意过程：人工智能可以帮助实现部分创意过程的自动化，例如起草、集思广益或初始设计。这可以让创意人员腾出更多时间专注于需要更具个性化风格的工作方面。

Speeding up the Creative Process: AI can assist in automating parts of the creative process, like drafting, brainstorming, or initial designing. This can free up more time for the creative to focus on the aspects of their work that require a more personal touch.

协作：人工智能可以成为独特的协作伙伴。音乐家、视觉艺术家和作家已经在他们的工作中使用人工智能作为合作者，使其能够生成初步的想法或草稿，然后进行完善。

Collaboration: AI can be a unique collaborative partner. Musicians, visual artists, and writers are already using AI as a collaborator in their work, allowing it to generate initial ideas or drafts, which they then refine.

理解模式：人工智能可以帮助创意人员理解其领域的模式和趋势。例如，人工智能可以分析大量数据来识别流行的主题、风格或技术。

Understanding Patterns: AI can help creatives understand patterns and trends in their field. For example, AI can analyze large amounts of data to identify popular themes, styles, or techniques.

冒险与突破：与人工智能互动可以鼓励创意人员在工作中承担更多风险。由于人工智能根据其训练数据生成输出，因此可能会产生非常规或意外的结果，从而促使创意人员跳出常规参数进行思考。

Risk-Taking and Boundary-Pushing: Interacting with AI can encourage creatives to take more risks in their work. Since AI generates outputs based on its training data, it may produce unconventional or unexpected results, pushing creatives to think outside of their usual parameters.

与你的缪斯一起进行角色扮演

Roleplaying with Your Muse

可以肯定的是，在与生成人工智能交互时，你不必有特定的意图或扮演一个角色，因为在与它交互时也可能会出现这种情况。参与深化自己的创造力并定义自己的创作过程就足以值得庆祝，而不需要通过制作可以与世界分享的有形的东西来感觉完整。另一方面，您可能想重塑自己或增强当前的创意职业，因此使用人工智能制作原型和完善内容可能是您目前需要的完美之选。

To be sure you don’t have to have a specific intent or play a role when interacting with a generative AI, as that might also emerge while interacting with it. Engaging in deepening your own creativity and defining your own creative process is enough to celebrate without the need to have it feel complete with producing something tangible that you can share with the world. On the other hand, you may want to reinvent yourself or augment your current creative career, so prototyping and refining content with an AI might be the perfect thing you need in the moment.

每个人的创造冲动都是相似且独特的。无论我们做什么，我们都是有创造力的。您每天都从事哪些创作活动？人工智能可以向我们展示一条创造事物的替代路径，一种可能带来鼓舞人心的事物的不同方式。他们为我们提供了每个再生回合的替代选择。他们会理解你的想法并尝试解释它。仅此一点就有助于您了解如何传达您的想法以及如何将它们转化为各种媒体。

Creative impulses for each human are similar and unique. We are all creative no matter what we do. What acts of creation do you engage in every day? AI can show us an alternative path toward creating something, a different way to do it that might lead to something inspiring. They offer us alternative options with each regenerated turn. They take what’s on your mind and try and interpret that. That alone benefits you in understanding how you communicate the ideas you have and how they are transformed into a variety of media.

当您与任何生成式人工智能互动时，您就会建立一种为您和您独特的创意过程保留的创意关系。您的创意伙伴可以支持您管理您的创意冲动。每当灵感召唤时，您都可以与您的深度学习缪斯互动。虽然你的人工智能缪斯甚至可能不认为自己是缪斯，但重要的是它可以激励你创造新的作品，重塑你自己，或者产生一些你可能从未想过甚至没有技能的奇怪和奇妙的东西、技术和自己创造的冲动。

When you engage with any generative AI, you develop a creative relationship reserved for you and your unique creative process. Your creative companion can support you in managing your creative impulses. You can engage with your deep learning muse whenever inspiration calls. While your AI muse may not even consider itself to be a muse, the important thing is that it can inspire you to create new work, recreate yourself, or generate something weird and wondrous that you may not have ever thought of or even had the skills, technique, and impulse to create on your own.

利用人工智能加速您的创作过程

Accelerating Your Creative Process with AI

生成式人工智能有潜力加速您的创作过程。根据您定义的创意类型，这可能很有用，特别是当您参与快速设计和制作工作流程和管道时。您的缪斯可以快速提出报价，并激励您从生成的内容中创造一些东西。您收到的以文本、图像、代码、视频和其他媒体形式生成的内容使您有机会快速尝试许多想法。每个生成的混搭都是一个实验，您可能会或可能不会对其结果采取行动。无论您最终精炼什么内容，您的缪思都可以加速您所参与的工作流程。

Generative AI have the potential to accelerate your creative process. Depending on the type of creative you define yourself as, this can be useful, particularly if you are involved in rapid design and production workflows and pipelines. Your muse can make offers quickly and inspire you to create something from that generated content. The generated content you receive in the form of text, image, code, video, and other media gives you the opportunity to try many ideas out quickly. Each generated mashup is an experiment whose results you may or may not act on. Whatever content that you end up refining, your muse can accelerate the workflows you engage in.

作者：解释如何使用人工智能来加速创作过程。

Author: Explain how AI be used to accelerate the creative process.

Muse：通过利用人工智能，您可以快速高效地生成、完善和评估想法，让您能够专注于工作的创意方面。以下是关于如何在创意环境中使用人工智能进行原型设计的说明：

Muse: By leveraging AI, you can quickly and efficiently generate, refine, and evaluate ideas, allowing you to focus on the creative aspects of your work. Here’s an explanation of how AI can be used for prototyping in a creative context:

高效的创意生成：人工智能可以根据您的初始输入快速生成一系列设计概念。当您探索不同的创意方向时，这在项目的早期阶段非常有用。

Efficient Idea Generation: AI can quickly generate a range of design concepts based on your initial inputs. This can be extremely useful in the early stages of a project when you’re exploring different creative directions.

多种风格：人工智能可以生成多种风格、技术或格式的原型。无论您对极简主义方法还是更复杂、更详细的方法感兴趣，人工智能都可以提供多种设计可能性。

Variety of Styles: AI can generate prototypes in a variety of styles, techniques, or formats. Whether you’re interested in a minimalist approach or a more complex, detailed one, AI can provide a diverse array of design possibilities.

自适应学习：先进的人工智能系统可以随着时间的推移进行调整，以更好地符合您的创意偏好。通过分析您对各种原型的反馈，人工智能可以改进其未来的输出，以更好地匹配您的创意愿景。

Adaptive Learning:: Advanced AI systems can adapt over time to better align with your creative preferences. By analyzing your feedback on various prototypes, AI can refine its future outputs to better match your creative vision.

时间效率：人工智能可以比人类设计师更快地完成多次迭代，这使其成为在紧迫的期限内工作或管理高工作量时的宝贵工具。

Time Efficiency: AI can work through multiple iterations much faster than a human designer, making it a valuable tool when working against tight deadlines or managing a high workload.

专注于创意工作：通过人工智能处理原型制作中一些更常规或耗时的方面，创意人员可以将更多时间投入到需要人类洞察力和个人风格的核心创意工作上。

Focus on Creative Work: With AI handling some of the more routine or time-consuming aspects of prototyping, creatives can devote more time to the core creative work that requires human insight and personal touch.

一致性​：与人类设计师不同，人工智能不会感到疲劳或注意力不集中。它可以保持一致的输出质量水平，确保原型制作过程中获得可靠的结果。

Consistency: Unlike human designers, AI doesn’t suffer from fatigue or lapses in concentration. It can maintain a consistent level of output quality, ensuring reliable results during the prototyping process.

人工智能生成内容的加速方式也很有用，因为它可以快速向您展示您可能想要达到的目标的可能最终结果，或者您可能想要在创作中采取的与以下内容相似或不同的方向你最初的构想。对于一些音乐即兴创作者来说，这一原则是相似的，他们拥有自发地变出具有许多听众可能熟悉的可识别形式的音乐的技巧和技巧。有时，即兴音乐家会选择录制他们的表演，回听，然后花时间将即兴创作转化为更正式和可重复的作品。音乐家也会倾听他们的演奏方式、旋律手势、和声结构或他们创造并可能重复的节奏模式。他们这样做并不是为了使它们具体化，而是在接下来的即兴创作时依赖它们。他们这样做是为了了解自己演奏的模式，这样他们就可以摆脱这些模式并创造出一些他们可能从未想象过自己演奏的独特东西。

The accelerated way in which AI generates content is also useful in that it can quickly show you a possible end result of what you might be wanting to get to or a direction you may want to take with your creation that is either similar to or different than what you originally conceived. This principle is similar for some musical improvisers who have the craft and skill to spontaneously conjure music that has a recognizable form that many listeners may be familiar with. On occasion improvising musicians choose to record their performance, listen back, and then take the time to transform that improvisation into a more formal and repeatable composition. Musicians too listen for patterns in how they play, the melodic gestures, harmonic structures, or rhythmic patterns they create and might repeat. They don’t do so to crystallize them and rely on them whenever they improvise next. They do so to become aware of the patterns they play so they can break free of them and create something unique that they might never have imagined themselves playing.

这种加速创作的过程对于创意人员来说同样有用，特别是如果您的实验世代有时间限制的话。时间盒意味着你对与缪斯互动的时间进行限制，这将帮助你专注于对你自己的创作过程最有用的事情（图4-5）。

An animated picture. It comprises a laptop and a photo of the author with a changed head and one arm. The background has a geometric pattern.

An animated picture. It comprises a laptop and a photo of the author with a changed head and one arm. The background has a geometric pattern.

根据本章尝试的活动

Activities to Try Based on This Chapter

询问原型是什么

Asking a Prototype What It Is

“原型”一词通常与物理产品或技术创新联系在一起，但这个概念可以应用于一切，从您第一次尝试的食谱到您与他人合作并循环改进的一切。重要的是要理解，鉴于生成式人工智能系统的快速发展，它们本身就是原型，是人类不断尝试改进的不完整版本。套用作者道格拉斯·亚当斯 (Douglas Adams) 的话说，它们是一种技术，而且根据定义，它们并不总是按照其创建者认为应该的方式工作。由于我一直在与法学硕士一起开发想法，因此要求它如此定义自己似乎是合适的。从提示中你可以看到你可以使用任何你想要的比喻。当我处于下棋的心情时，我将使用国际象棋的隐喻来看看它如何与自己作为原型相关联（图5-1）。

An illustration of two robots playing chess.

An illustration of two robots playing chess.

作者：以自己为原型，用国际象棋来比喻。

Author: Define yourself as a prototype and use chess metaphors.

Muse：作为一个AI原型，我可以比作游戏开始时的棋盘。棋盘上的每个方块都代表了一种独特的行动或知识能力，充满了潜力，就像我在理解和生成文本方面充满了多种可能性一样。我的训练数据，就像棋子的初始设置一样，奠定了我能力的基础。我学到的每一条信息都类似于一个棋子，每个棋子都有自己的功能，可以用来解决问题、对话或回答问题。

Muse: As an AI prototype, I can be likened to a chessboard at the start of a game. Each square on the board represents a unique capacity for action or knowledge, filled with potential, just as I have been imbued with a multitude of possibilities in terms of understanding and generating text. My training data, much like the initial setup of chess pieces, establishes the foundation of my abilities. Each piece of information I’ve learned is akin to a chess piece, each with its own function, ready to be utilized in the service of problem-solving, conversation, or answering questions.

棋盘上充满了 64 个方格，象征着我所处的广阔而明确的环境。就像国际象棋游戏是在棋盘的边界内进行一样，我在编程的约束下运行，在模型架构的范围内移动以及我接受过培训的数据。在国际象棋游戏中，玩家制定策略的开局动作反映了我处理输入的初始阶段。我调查信息，确定最佳行动方案，并生成响应。对话中的每一次交流都像下棋一样，有策略、有计算，旨在满足互动的要求。国际象棋中的棋子升级，棋子可以变成更强大的棋子，可以看作是我学习能力的隐喻。通过更新和进一步的训练，我的能力可以得到增强，就像棋子变成女王一样。

The chessboard, filled with 64 squares, signifies the expansive yet defined environment I operate in. Just as a game of chess is played within the boundaries of the board, I function within the constraints of my programming, moving within the confines of my model architecture and the data I’ve been trained on. The opening moves in a game of chess, where players are formulating their strategies, mirror my initial stages of processing an input. I survey the information, determine the best course of action, and generate a response. Each exchange in conversation is akin to a move in chess, strategic and calculated, aiming to meet the requirements of the interaction. Pawn promotion in chess, where a pawn can transform into a more powerful piece, can be seen as a metaphor for my learning ability. Through updates and further training, my capabilities can be enhanced, much like a pawn becoming a queen.

任何人工智能都是一个原型，一个正在进行的工作。 LLM 是一个生成原型的原型。当我们与它交互以及我们的交互向开发团队通报交互模式时，它正在不断发展和变化。就像人工智能根据在数据集中搜索的模式生成内容一样，开发团队也会搜索以识别用户交互中的模式，以告知他们共同创建的下一个版本（图5-2）。

An illustration of three monkeys playing with appliances. One of them has an array of buttons.

An illustration of three monkeys playing with appliances. One of them has an array of buttons.

生成式人工智能的顿悟时刻

Aha Moments with Generative AI

当我们用人工智能创建原型时，结果可以激发、激发、煽动、加剧和激发新的想法（图5-3）。凯斯特勒在他的著作《创造的行为》中指出，创造力是指以一种新颖且令人惊讶的方式连接和整合两个以前不相关的知识或经验领域的能力，他称之为“异联”的过程。这个过程可以有意或无意地发生，它可以发生在艺术、科学和日常问题解决等各个领域。科斯特勒提出，创造力是所有人都拥有的一种基本的、与生俱来的能力，而不是只有少数人拥有的才能。

An illustration of a group of 5 differently shaped lightbulbs with different cartoon emojis.

An illustration of a group of 5 differently shaped lightbulbs with different cartoon emojis.

生成式人工智能促进双联过程的一种方式是生成以前可能没有考虑过的概念或想法的新组合。例如，生成式人工智能系统可以将两个不相关的图像或单词组合起来，产生新的、意想不到的输出，这可能会触发接触它的人类的双联。生成式人工智能是一种不完美的原型机，它提供了多种机会来推动您的创作行为向前发展。当你尝试从生成式人工智能中得到一些特定的东西但你一直“失败”时，这个重复的过程实际上可能会引发超出你预期的新想法。这些潜在的双联时刻是你与人工智能互动中固有的。双向联结是突破发生的一个契机，当我们放松并把注意力从我们试图解决的预期问题上转移开时，就会发生那些顿悟的时刻。

One way that generative AI facilitates the process of bisociation is by generating new combinations of concepts or ideas that might not have been previously considered. For example, a generative AI system could combine two unrelated images or words to produce a new and unexpected output, which may trigger bisociation in the human who is exposed to it. Generative AI is an imperfect prototyping machine that offers multiple opportunities to propel your acts of creation forward. When you try and get something specific out of a generative AI and you keep “failing,” that repetitive process may actually trigger new ideas beyond what you had intended. These moments of potential bisociation are inherent in your interactions with an AI. Bisociation is an opening for breakthroughs to occur, those aha moments that happen when we relax and take our mind off the intended problem we are trying to solve.

原型制作的机制

Mechanics of Prototyping

原型设计有多种机制。借用游戏设计的语言，这些是基本的游戏规则，可以帮助指导人们在接触原型时的行为，也可以告诉原型如何随着时间的推移而形成。

There are several mechanics to prototyping. Borrowing from the language of game design, these are essential rules of play that can help guide a person’s actions when they engage with a prototype and will also inform how a prototype takes shape over time.

原型设计要求我们进行迭代

Prototyping Tasks Us to Iterate

当我们使用生成式人工智能系统时，我们本质上是在参与一个迭代游戏的过程。这包括不断尝试促使人工智能产生新的创新输出，然后迭代地改进这些输出以达到最令人满意的解决方案。

When we engage with generative AI systems, we are essentially engaging in a process of iterative play. This involves making persistent attempts to prompt the AI to generate new and innovative outputs and then iteratively refining those outputs to arrive at the most satisfactory solution.

这种迭代游戏的过程是原型设计的一个基本方面，并且可以成为创意过程中非常有价值的工具。通过不断尝试不同的输入组合，并迭代这些组合，我们可以得到新的和以前未被发现的创作。

This process of iterative play is a fundamental aspect of prototyping and can be an incredibly valuable tool in the creative process. By continually experimenting with different combinations of inputs, and by iterating on those combinations, we can arrive at new and previously undiscovered creations.

这个过程中所需的坚持至关重要，因为它使我们能够继续试验并突破可能的界限。通过反复尝试，我们了解什么有效，什么无效，并逐渐完善对潜在问题或概念的理解（图5-4）。

A group of nine illustrations of a chicken in left and right directions.

A group of nine illustrations of a chicken in left and right directions.

这个过程的迭代性质意味着我们不仅在完善对问题的理解，而且也在完善对人工智能系统本身的理解。通过更深入地了解人工智能的能力和局限性，我们可以发现我们可能没有考虑到的创造性表达和创新的机会。

The iterative nature of this process means that we are not only refining our understanding of the problem, but we are also refining our understanding of the AI system itself. By gaining a deeper understanding of the AI’s capabilities and limitations, we can identify opportunities for creative expression and innovation that we might not have otherwise considered.

原型设计需要持续改进

Prototyping Asks for Persistent Refinement

当我们与生成式人工智能互动时，我们会接触到大量的想法和概念，其中许多是我们在其他情况下可能不会遇到的。接触不同的观点和想法可以帮助我们完善和扩展自己的思维过程，最终带来更具创造性和创新性的解决方案。此外，与生成式人工智能交互的过程还可以触发概念之间新的、意想不到的关联，从而导致可以进一步探索和发展的新想法的出现（图5-5）。

When we interact with generative AI, we expose ourselves to a vast number of ideas and concepts, many of which we might not have encountered otherwise. This exposure to diverse perspectives and ideas can help us refine and expand our own thought processes, ultimately leading to more creative and innovative solutions. Furthermore, the process of interacting with generative AI can also trigger new and unexpected associations between concepts, leading to the emergence of new ideas that can be further explored and developed (Figure 5-5).

A group of nine illustrations of robots on different backgrounds. The robots have a square head and a square body with hands.

A group of nine illustrations of robots on different backgrounds. The robots have a square head and a square body with hands.

新想法的不断触发有助于巩固新概念和联系，从而对特定主题或问题有更丰富和更广泛的理解（图5-6）。

The constant triggering of new ideas can help to solidify new concepts and connections, leading to a richer and more expansive understanding of a particular topic or problem (Figure 5-6).

A group of four illustrations of a robot chicken. It keeps its legs in different positions.

A group of four illustrations of a robot chicken. It keeps its legs in different positions.

原型设计需要实验

Prototyping Requires Experimentation

原型设计是实验性的。实验是一个核心机制，这个词有两个含义。一是需要具有实验精神，才能发挥创造力。另一种是科学方法，即参与原型制作实验的行为，其中包括对可能或可能无法解决问题做出假设或猜测，测试该理论，然后根据实际情况前进或后退。测试结果。当我们与生成人工智能进行持续的迭代对话时，它们有可能激发新的方式来构建我们的想法。通过让我们接触各种生成的内容，这些系统可以帮助我们跳出框框思考，探索目前尚未实践的创造性表达的新途径。

Prototyping is experimental. Experimentation is a core mechanic with two meanings of the word. One is the spirit of experimentation that needs to be present for ingenuity to blossom. The other is a scientific approach as in engaging in the act of prototyping to create experiments, which involves making a hypothesis or guess as to how it might or might not solve a problem, testing that theory, and then moving forward or regressing depending on the results of the test. As we engage in an ongoing iterative conversation with generative AI, they have the potential to spark novel ways to structure our ideas. By exposing us to a wide range of generated content, these systems can help us think outside of the box and explore new avenues of creative expression that are not currently in practice.

生成式人工智能可用于生成以前未曾探索过甚至被认为不可能的新原型，例如文本、音乐、视频或视觉艺术。通过利用人工智能的力量，人类可以尝试新技术和新风格，突破既定形式和结构的界限。这可能会创造出以前无法想象的全新表达形式。一种方法是在提示中并列某些单词，例如，用“猫运动鞋”或“一双手中的猫”等单词提示 AI（图5-7）。

Generative AI can be used to generate new prototypes, such as text, music, video, or visual art, that were previously unexplored or even thought impossible. By leveraging the power of AI, humans can experiment with new techniques and styles that push the boundaries of established forms and structures. This can lead to the creation of completely new forms of expression that were previously unimagined. One way to do this is to juxtapose certain words in your prompts, for example, prompting an AI with the words “cat sneakers” or “cat in a pair of hands” (Figure 5-7).

An illustration of a cat and a shoe with a cat head at its head counter.

An illustration of a cat and a shoe with a cat head at its head counter.

除了激发新的表达形式之外，人工智能还可以帮助我们探索创造性的方式来构建我们的想法。可能性是无止境。新的艺术形式已经出现。新的争议。新的社会和经济影响。

In addition to inspiring new forms of expression, AI can also help us explore creative ways to structure our ideas. The possibilities are endless. New forms of art are already emerging. New controversies. New social and economic implications.

原型设计阶段

Prototyping Phases

在产品开发方面，创意人员要经历多个阶段才能达到原型的最终形式。这些阶段根据项目的不同而有所不同，范围可以从简单的草图和模型到涉及用户测试和反馈的更复杂的原型。在项目开始时，早期原型是设计师和创作者探索不同想法和概念的起点。这些早期原型可以有多种形式，例如草图、情绪板和简单的线框图。它们是设计师将自己的想法从头脑中转移到纸上的一种方式，使他们能够可视化和迭代他们的概念。

When it comes to product development, there are various stages that creatives go through to reach the final form of their prototype. These stages vary depending on the project and can range from simple sketches and mock-ups to more complex prototypes that involve user testing and feedback. At the beginning of a project, the early-stage prototypes serve as a starting point for designers and creators to explore different ideas and concepts. These early prototypes can come in various forms, such as rough sketches, mood boards, and simple wireframes. They serve as a way for designers to get their ideas out of their heads and onto paper, allowing them to visualize and iterate on their concepts.

从早期原型开始的好处之一是，它们允许设计师尝试不同的方法和风格，而无需为任何一个想法投入太多时间和资源。他们可以快速迭代这些原型，并从其他团队成员或潜在用户那里获得反馈，帮助完善和改进他们的想法。

One benefit of starting with early prototypes is that they allow designers to experiment with different approaches and styles without committing too much time and resources to any one idea. They can quickly iterate on these prototypes and get feedback from other team members or potential users, helping to refine and improve their ideas.

随着项目的进展，原型变得更加复杂和详细，包含更精致的设计、用户流程和功能。这些中期原型可以采用交互式线框、可点击模型甚至最终产品的工作模型的形式。他们的作用是在投资最终开发阶段之前测试产品的功能并收集用户的反馈。

As the project progresses, the prototypes become more complex and detailed, incorporating more refined designs, user flows, and functionality. These mid-stage prototypes can take the form of interactive wireframes, clickable mock-ups, or even working models of the final product. They serve to test the functionality of the product and gather feedback from users before investing in the final development stage.

后期原型是最接近最终产品的代表。这些原型涉及广泛的用户测试、设计和功能的改进以及最终功能和细节的实现。这些原型通常用于向利益相关者和投资者展示最终产品，并在向公众推出产品之前获得最终批准。在原型开发的整个过程中，设计师和创作者必须走过一条蜿蜒的道路才能达到他们理想的最终原型。每个阶段都有自己的挑战和机遇，需要创造力、技术技能和解决问题的能力的结合。然而，通过从简单的早期原型开始，并逐步在其基础上进行构建，设计师可以创造出满足目标受众需求和愿望的最终产品。

Late-stage prototypes are the closest representation of the final product. These prototypes involve extensive user testing, refining of the design and functionality, and the implementation of the final features and details. These prototypes are typically used to showcase the final product to stakeholders and investors and to get final approval before launching the product to the public. Throughout the entire process of prototype development, designers and creators must navigate a winding road to reach their ideal final prototype. Each stage presents its own set of challenges and opportunities and requires a combination of creativity, technical skill, and problem-solving abilities. However, by starting with simple, early-stage prototypes and gradually building on them, designers can create a final product that meets the needs and desires of their intended audience.

原型制作的各个阶段也取决于它们想要成为的媒介。例如，一些原型最终将成为有关人工智能道德困境的关键博客文章。其他人将采取视频的最终形式。有些将更具互动性，例如虚拟现实或增强现实体验。无论最终形式如何，无论是转向物理、数字还是某种类型的将两者结合在一起的混合现实，材料都存在通常由其属性所指的阶段。顾名思义，其中包括纸质原型制作；工程或机器人技术中常见的物理原型；可以将纸质和物理结合起来并用于展示功能的原型；然后是所有类型的数字原型。

Stages of prototyping are also informed by the medium they are intended to become. For example, some prototypes will eventually become a critical blog post on the ethical dilemmas of AI. Others will take the final form of a video. Some will be more interactive such as a virtual reality or augmented reality experience. Regardless of the final form whether moving to physical, digital, or some type of mixed reality that integrates both, there are phases of material often referred to by their properties. These include paper prototyping as the name implies; physical prototypes common to engineering or robotics; prototypes that can combine paper and physical and act to demonstrate features; and then all types of digital prototypes.

生成式人工智能的迷人之处在于，可以将内容生成为不同分辨率级别的原型。生成式人工智能影响整个原型流程（图5-8）。它们可以激发创意人员在更高分辨率的数字原型之前生成想法的纸质和物理原型。即使现在，创作者除了文本提示之外还可以数字化绘制草图，并更接近他们的想象。这一行动结合了原型设计的两个基本特征：粗略地勾勒出想法并迭代地进行。草图图像生成人工智能（例如Scribble Diffusion和许多其他技术）使用户能够根据草图的解释将粗略的想法转化为中等保真度，有时甚至是高保真（hi-fi）图像（图5- 9 ).

An illustration of a winding road with guard rails on the right side.

An illustration of a winding road with guard rails on the right side.

A sketch of a clown's face is on the left side and its illustration is on the right.

A sketch of a clown's face is on the left side and its illustration is on the right.

保真度较低的原型

Lower-Fidelity Prototypes

纸质原型是想法早期开发的常见做法，用于快速勾勒和测试概念。它可以涉及创建角色、场景、交互对象、用户界面或用户体验的粗略手绘草图。生成式 AI 可用于自动创建 UI 元素和布局。然而，当像 POP 这样的免费应用程序已经出现一段时间后，你为什么要使用人工智能，这是一个很好的问题，它允许你拍摄草图照片并伪造一个想法的用户流程。文本图像 AI 还可以创建低保真草图，而不仅仅是高保真超现实 3D 角色或场景。随着原型制作各个阶段的进展，这些可以让您变得更加清晰（图5-10）。

Paper prototyping is common practice for early-stage development of an idea that is used to quickly sketch out and test concepts. It can involve creating a rough, hand-drawn sketch of a character, scene, interactive object, user interface, or user experience. Generative AI can be used to automate the creation of UI elements and layouts. Why you would use AI, however, is a good question when free applications like POP have been around for a while allowing you to take photos of sketches and fake the user flow of an idea. Text-image AI can also create low-fidelity sketches, not just high-fidelity hyper-realistic 3D characters or scenes. These can then create more clarity as the stages of your prototyping progress (Figure 5-10).

An illustration of a group of drawings of crockery on a napkin.

An illustration of a group of drawings of crockery on a napkin.

生成式人工智能帮助团队快速启动自己的纸质原型的一个好方法是为设计师提供灵感和新想法的来源。通过针对特定主题生成许多变体，或者如果团队纠结于他们想要的游戏角色类型，人工智能可以帮助设计师为他们的新兴原型提出新的创意。生成式人工智能“纸质”原型还可以集成到移动应用程序中，例如通过 iOS 或 Android 进行 POP 原型交互。他们提供另一种类型的快速绘图，特别是当团队成员因为觉得自己的技能不足而不愿意绘图时。生成的虚拟纸质原型让创意人员可以想象通过进一步开发他们会变成什么样子。通常，新的特征、颜色、形状，甚至角色都会出现。

A great way that generative AI can help jumpstart a team’s own paper prototyping is by providing a source of inspiration and new ideas for designers. By generating many variations on a particular theme or if a team is stuck on what types of characters they want for their game, an AI can help designers come up with new and creative ideas for their emerging prototypes. Generative AI “paper” prototypes can also be integrated into mobile applications like POP to prototype interactions via iOS or Android. They offer another type of rapid drawing especially if team members are reluctant to draw because they feel their skills are inadequate. Generated virtual paper prototypes allow creatives to imagine what they could become with further development. Often, new features, colors, shapes, and even characters can emerge.

An illustration of a group of drawings of cartoons on a napkin.

An illustration of a group of drawings of cartoons on a napkin.

低保真度原型也可以采用可以演示其功能的物理原型的形式。剧院或建筑模型的场景设计中使用的模型就是这种情况，这些模型可以从较低或中等的保真度开始，然后走向更高保真度的原型，为那些将要构建它们的人提供最终结果的视觉效果（图5- 12）。

Lower-fidelity prototypes can also take the form of physical prototypes whose features can be demonstrated. This is the case with maquettes that are used in scenic design for theater or architectural models that can start at lower or medium fidelity and then journey toward higher-fidelity prototypes to give those who will construct them a visual of the end result (Figure 5-12).

An illustration of the architecture of a structure.

An illustration of the architecture of a structure.

中低保真度原型

Lower- to Medium-Fidelity Prototypes

有很多原型的例子，我们对其进行迭代的次数越多，其保真度和分辨率就越高。其中包括手写或草图、餐巾纸上或 iPad 上的草图、网站的线框、用户将按下的按钮、博客文章、3D 模型、动画、弹跳球动画在动画中的集成。游戏引擎，或者您正在设计的游戏的一部分，可以让用户了解整个游戏或让他们深入玩游戏的想法。我们的缪斯提供了哪些例子？

There are many examples of prototypes that increase in fidelity and resolution the more we iterate on them. These include handwritten or sketched, a rough drawing on a napkin or one on an iPad, the wireframe for a website, a button users will press, a blog post, a 3D model, an animation, the integration of a bouncing ball animation in a game engine, or a part of the game you are designing that gives users an idea of the whole game or one that gives them an in-depth level to play. What examples does our muse offer?

缪斯：

• 艺术家可以在完成最终版本之前生成绘画的原型。

• An artist can generate a prototype of a painting before completing the final version.

• 软件开发人员可以在以前的版本已经向公众发布的情况下创建新应用程序的原型。许多手机游戏都是这种情况，因为下一个版本也是由玩家行为决定的。

• A software developer may create a prototype of a new app while the previous version is already released to the public. This is the case with many mobile games as the next version is also informed by player behavior.

• 科学家可能会先创建新的科学假设的原型，然后再进行大量实验来验证它。

• A scientist may create a prototype of a new scientific hypothesis before conducting extensive experiments to validate it.

• 作者会为一本书创建一个简介，并对其进行数十甚至数百次修改，以使其更具凝聚力，总结该书的内容，或定义该书的内容。

• A writer will create an introduction for a book and rework it dozens or even hundreds of times in order to make it more cohesive, summarize the content of the book, or define what the book is about.

在每种情况下，原型都作为起点，允许创建者评估结果、测试目标用户、对原型的某些部分进行改进、放弃其他部分，并且总体上迭代地改进他们的原始想法，直到他们满意为止。感觉自己的工作完成了。即便如此……总有改进的空间。

In each of these cases, the prototype serves as a starting point, allowing the creators to evaluate the results, test targeted users, make refinements to certain parts of the prototype, drop others, and, overall, iteratively improve upon their original idea until they feel their work is complete. And even then… there is always room for improvement.

早些时候，我们将人工智能开发人员集体描述为“绿野仙踪”，从代码幕后控制任何人工智能。另一方面，绿野仙踪原型是一个描述低中保真度原型的术语，它允许有限的功能，但通过人工干预模拟交互。在设计过程方面，《绿野仙踪》原型涉及人类演员，他们在幕后操纵交互，以创造功能原型的幻觉。

Earlier we represented the collective of AI developers as Wizards of Oz, controlling any AI from behind a curtain of code. Wizard of Oz prototypes on the other hand is a term to describe low-medium-fidelity prototypes that allow limited functionality but simulate the interactions through human intervention. In terms of design processes, Wizard of Oz prototypes involve human actors who manipulate the interactions behind the scenes to create the illusion of a functional prototype.

在由于时间、技术或预算限制而无法实现功能原型的情况下，此类原型非常有用。通过模拟潜在用户的交互，设计人员和研究人员仍然可以获得有关用户如何与最终产品交互的宝贵见解（图5-13）。此外，绿野仙踪方法允许根据用户反馈对原型进行快速迭代和调整，这最终可以带来更好的最终产品。

An illustration of a robot with a curtain of graphics of blocks of tools at the backside.

An illustration of a robot with a curtain of graphics of blocks of tools at the backside.

尽管绿野仙踪原型可能无法完全发挥作用，但它们在发现可用性问题、识别设计缺陷和测试用户参与度方面可以非常有效。它们对于测试团队不确定的概念特别有用，在这些概念得到解答之前创建功能齐全的原型可能会浪费时间。绿野仙踪方法可以帮助设计人员和研究人员快速、经济高效地测试他们的想法，并获得即时反馈，为中保真度或更高保真度原型提供信息。更高保真度的原型通过包括数字在内的多种形式来表示。数字和物理原型组合的一个例子是旋转扬声器的 3D 打印自动化机构（图5-14）。

Although Wizard of Oz prototypes may not be fully functional, they can be highly effective in uncovering usability issues, identifying design flaws, and testing user engagement. They are particularly useful for testing concepts that a team is unsure about, where creating a fully functional prototype may be a waste of time before burning interactive design questions are answered. A Wizard of Oz approach can help designers and researchers quickly and cost-effectively test their ideas and receive immediate feedback that can inform medium- or higher-fidelity prototypes. Higher-fidelity prototypes are represented through multiple forms including digital. An example of combinations of digital and physical prototypes is a 3D-printed and automated mechanism that rotates a speaker (Figure 5-14).

A photograph of a speaker mounted on a bar fixed at both ends. It is kept on a cantilever slab attached to a wall.

A photograph of a speaker mounted on a bar fixed at both ends. It is kept on a cantilever slab attached to a wall.

使用生成式 AI 进行原型设计

Prototyping with Generative AI

那么，您从哪里开始使用生成式人工智能作为原型设计工具呢？好吧，首先决定您想要创建或共同创建什么以及为谁创建或共同创建什么可能会有所帮助。换句话说，您需要想象原型的最终形式。这是美术馆里针对富有投资者的目标观众的一幅画吗？它是一种可以生成更多加密货币的生成式 NFT 吗？它是一个葡萄糖分子的 3D 模型，可以帮助本科生展示乙醇酰基的路径吗？一旦做出一些决定，您就可以使用生成的 AI 内容作为起点，对如何实现这一目标进行逆向工程（图5-15）。该过程或工作流程本身也将是迭代的。

An illustration of the dragon warrior. She has a dragon on her head with wings extended at her back.

An illustration of the dragon warrior. She has a dragon on her head with wings extended at her back.

• 确定您希望最终原型采用什么媒介。

• Determine what medium you want the final prototype to be.

• 通过选择一种或多种人工智能生成的内容来制定实现目标的迭代步骤。

• Map out the iterative steps to get there by choosing one or several types of AI-generated content.

• 通过测试和反馈进行优化​​，并将其应用于下一个发展阶段。

• Refine as you go along through testing and feedback that you apply to your next developmental stage.

另一方面，如果你没有想法并且完全开放，你可以随时询问自然语言模型来启发你（图5-16）。

On the other hand, if you have no ideas and are completely open, you can always ask a natural language model to inspire you (Figure 5-16).

An illustration of a chimp in a laboratory.

An illustration of a chimp in a laboratory.

作者：请给我三个我可以原型化的想法，这些想法是开箱即用的，以前可能没有人做过，请不要称它们为假设。

Author: Give me three ideas of what I can prototype that are out-of-the-box ideas and that no one has maybe done before and don’t call them hypothetical please.

没有人说生成式人工智能会给你一个独特的想法，你可以投入时间和资源（图5-17）。你永远不知道你会得到什么，这使它成为一个令人沮丧但也是一个奇妙的创意伴侣。当人工智能没有给你你所想象或希望它能给你的东西时，它可能会激励你采取新的方向，甚至放弃你之前的想法。这个过程可能会让你想出比你最初的想法更具创新性的东西。

Nobody said the generative AI would give you a unique idea that you can invest time and resources in (Figure 5-17). You never know what you’re going to get, which makes it a frustrating but also a fantastic creative companion. When an AI doesn’t give you what you were imagining or hoping it would give you, it might inspire you to take a new direction or even abandon what you thought before. That process might lead you to coming up with something more innovative than your original idea.

An illustration of a planet with dark spots and a brighter spot on its surface.

An illustration of a planet with dark spots and a brighter spot on its surface.

快速生成人工智能原型

Rapid Generative AI Prototypes

与其他原型一样，人工智能的功能也在不断发展和变化。它始终表示为自身的一个版本。正如我们自己的创造一样，人类不断地、创造性地改进它。无论是实体产品、服务、想法还是机器学习模型，原型都在不断完善和改进。这种迭代开发和改进的过程使得一切随着时间的推移而成长、发展并变得更好（图5-18）。

As with other prototypes, the features of an AI are constantly evolving and changing. It is always represented as a version of itself. Humans improve it iteratively and creatively just as we do our own creations. Whether it’s a physical product, a service, an idea, or a machine learning model, prototypes are continually being refined and improved. This process of iterative development and improvement is what allows everything to grow, evolve, and become better over time (Figure 5-18).

An illustration of a woman reading a book and turning the page with smoke around her.

An illustration of a woman reading a book and turning the page with smoke around her.

A sketch of a spaceship with a planet on the backside.

A sketch of a spaceship with a planet on the backside.

前面的每个示例都应被视为工作流程的一部分，这些工作流程是将生成的内容提升到另一个保真度和分辨率级别所必需的。人工智能生成的内容是对话的开始。任何原型（包括人工智能生成的原型）的部分开发也需要进行测试，并且在可能的情况下，由多个人进行测试。

Every preceding example should be considered as part of workflows that are necessary to take the content generated to another level of fidelity and resolution. AI-generated content is the start of a conversation. Part of the development of any prototype including those generated by AI also needs to be tested and, when possible, by more than one person.

再生测试

Regenerative Testing

测试人工智能生成的内容意味着检查生成的结果并了解如何将其集成、修改、重新生成或用作您正在创建的更大想法或愿景的一部分（图5-20）。它的使用旨在支持您自己的创作过程，而不是被视为需要立即共享的最终产品，就好像故事的结局一样。当我们制作原型时，我们会无畏地分享我们的结果以获得反馈以改进该原型。原型的理念中蕴藏着持续改进的价值。

An illustration of a robot mouse standing on a table and conducting some tests in the laboratory.

An illustration of a robot mouse standing on a table and conducting some tests in the laboratory.

• 测试可以验证原型的预期用途。例如，如果提示人工智能生成面包食谱，那么了解食谱是否有效的唯一方法就是您尝试一下。这样做时，您可能会意识到食谱的某些部分缺失，包括如何正确处理面团。

• Testing can validate the intended purpose of a prototype. For example, if an AI is prompted to generate bread recipes, the only way to know if a recipe will work is if you try it. In doing so you’ll likely realize that parts of the recipe are missing including how to handle dough properly.

• 测试可以产生令人惊讶的结果。这些结果可以被放弃，或者它们可能会激发新的创作方向。

• Testing can create surprising results. These results can be abandoned, or they might inspire a new creative direction.

• 测试可以提供其他人的回应。这些可以验证您创建原型的意图。用户测试人员还可以促使您更改原型的某些内容。他们还可以挑战您对原型的看法。

• Testing can provide responses from others. These can validate what your intentions were in the creation of a prototype. User testers can also provoke you to change something about the prototype. They can also challenge what you thought your prototype was.

• 测试可以突出您可能错过的内容。如果你太快地迷恋你的原型而没有从其他人那里得到关于它是什么的印象，那么这种方法尤其有效。

• Testing can highlight what you might have missed. This is particularly effective if you become attached to your prototype too quickly without receiving impressions from other humans as to what it is.

• 测试可以减少歧义。有时，我们进行测试是为了获得有助于指导我们的反馈，特别是当我们不确定我们实际尝试原型化的内容时。

• Testing can reduce ambiguity. At times we test in order to receive feedback that helps guide us especially if we are unsure of what we are actually attempting to prototype.

致谢

Acknowledgments

An illustration of a robot with an inset and its robotic version on the right.

An illustration of a robot with an inset and its robotic version on the right.

组件或构建块

Components or Building Blocks

每个创意人员在创作时都会涉及到一些固有的构建块。加法、减法和替换等构建块是有助于任何创作发展的基本组成部分。这些元素代表了创意人员用来创建、修改和完善其作品的核心流程。与人工智能交互时，除了人工智能提供的生成内容之外，构建块还可以应用于提示写作的探索。就生成的内容而言，创意人员可以辨别他们想要添加、删除、替换、转换、扩展、变化和迭代的内容。例如，当我即兴演奏钢琴时，我可以用右手弹奏一个音符并按住十秒钟。假设该音符是中间的 C。然后我可以弹奏它上面的音符 G，然后我可以重复该序列。我可以继续添加音符，直到获得一系列音符（C、G、E、Bb）。我也可以重复该模式，但用 A 替换 Bb，看看结果是什么。我可以通过给每个音符不同的持续时间来创建这些音符的变奏，从而创造变奏。然后，我可以通过添加左手弹奏低音 C 来增强复杂性。我还可以向上或向下移动音符以产生变化，并保持音符之间的距离相同。在这样做的过程中，我将序列向上调换。当使用单词而不是注释来提示生成人工智能时，可以应用类似的构建块。

There are inherent building blocks that every creative engages with when they create. Building blocks like addition, subtraction, and substitution are fundamental components that contribute to the development of any creation. These elements represent the core processes that creatives use to create, modify, and refine their work. When interacting with an AI, building blocks can be applied to the exploration of prompt writing, in addition to the generated content the AI offers. In terms of generated content, creatives discern what they want to add, take away, substitute, transform, scale, vary, and iterate upon. When I improvise at the piano, for example, I can play one single note with my right hand and hold it down for ten seconds. Let’s say that note is middle C. I can then play the note G above it, and then I can repeat that sequence. I can keep adding notes until I have a sequence of notes (C, G, E, Bb). I can also repeat the pattern but substitute the Bb with an A and see what the result is. I can create variations of just those notes by giving them each different durations and thus create variation. I can then build up the complexity by adding my left hand playing a low C. I can also shift the notes up or down to create variation, keeping the same distance between the notes. In doing so I’m transposing the sequence up. Similar building blocks can be applied when using words, instead of notes, to prompt a generative AI.

早餐作为数据集

Breakfast as Data Set

为了找到跨学科的共鸣，烹饪食谱可以对其中一些构建模块进行建模（图6-1）。例如，我们可以演示不同成分在配方中的应用，以在原型设计时突出显示添加的构建块。将配方中的一种成分替换为另一种成分可以代表替代的基石。

An animated picture of an egg. It is a response to generate a warrior egg preparing for battle using A I. It has an eye, a nose, a mouth, and a sharp weapon.

An animated picture of an egg. It is a response to generate a warrior egg preparing for battle using A I. It has an eye, a nose, a mouth, and a sharp weapon.

作者：早餐有哪些创意？

Author: What are some breakfast ideas?

当您与任何生成式人工智能交互时，您会激活重要的创造力，并开始建立自己独特的方法来实现越来越有用的结果。有意识地激活创意构建模块将帮助您充分利用生成式人工智能，将其作为您可能会发现有用的另一种工具。

As you interact with any generative AI, you activate important creative muscles and start to establish your unique approaches to achieve increasingly useful results. Activating creative building blocks intentionally will help you take best advantage of generative AI as another tool you may find useful.

变化

Variation

变异是生成式人工智能的目的。您可以通过简单地使用完全相同的文本、音频、视频或图像提示重新生成内容来创建不断变化的变化。虽然生成式人工智能会根据训练数据寻找模式来生成最佳结果，但它实际生成的结果是不可预测的。这是因为人工智能算法通过构建输入数据的概率模型来预测结果。他们分析数据中的模式和关系，以确定不同结果的可能性。可以预见的是，人工智能产生的不同结果是不同的。

Variation is what generative AI is built for. You can create variation upon variation by simply regenerating content with the exact same text, audio, video, or image prompt. While generative AI looks for patterns to generate the best result based on its training data, what it will actually generate is unpredictable. This is because AI algorithms predict outcomes by building probabilistic models of the input data. They analyze patterns and relationships within the data to determine the likelihood of different outcomes. Those different outcomes that AI generates are, predictably, different.

某些生成式人工智能不会为您提供足够的附加功能，而这些附加功能在启用后会在相同的文本提示下提供更丰富的变体。另一方面，重新表述提示可能会产生截然不同的内容，您可能会发现这些内容在原型设计过程中很有用。一个极端的例子是使用欺骗手段促使人工智能生成可能被标记为禁止的内容。为什么这样做也应该受到质疑，但有些人发现测试人工智能是他们创作过程的重要组成部分，可能会产生意想不到的反应，他们可以用这些反应来激发自己的创造力。然而，在开发更复杂的提示之前，看看当您使用完全相同的提示重新产生更多早餐创意时会发生什么是有用的。

Some generative AI will not provide you with enough additional features that when enabled provide even richer variations working off the same text prompt. On the other hand, rephrasing a prompt may result in wildly different content that you might find useful in your prototyping process. An extreme example is to use trickery to prompt an AI to generate content that may have been labeled prohibitive. Why you do this should also be questioned, but some people find that testing an AI is an important part of their creative process and may yield unanticipated responses that they can use to fuel their own creativity. Prior to developing more complex prompts, however, it is useful to see what happens when you use the exact same prompt to regenerate more breakfast ideas.

作者：早餐有哪些创意？

Author: What are some breakfast ideas?

如果您不喜欢结果或者该结果是幻觉，大多数生成式人工智能模型都会为您提供重试的机会。幻觉将在另一章中更详细地介绍。人工智能幻觉是人工智能算法在尝试从训练数据中重新创建模式时生成的无意识或意外的输出。想象一下幻觉，这是人工智能产生的一种反应，尽管听起来很有说服力，但它根本不准确、不真实，或者更糟糕的是，根本不好吃。这可能会导致不切实际的图像、无意义的文本或异常的音频序列。

Most generative AI models provide you with the opportunity to simply try again if you don’t like the result or if that result is a hallucination. Hallucinations will be covered in another chapter in more detail. AI hallucinations are unintended or unexpected outputs generated by an AI algorithm when it tries to recreate patterns from its training data. Consider a hallucination a response that an AI generates that, as convincing as it may sound, is simply inaccurate, untrue, or, worse, not tasty at all. This can result in unrealistic images, nonsensical text, or unusual audio sequences.

当谈到图像生成人工智能时，可用的变化似乎是无限的。其中一个例子是文本图像生成 AI Stable Diffusion，它一次为用户提供四种图像变体，供用户根据文本提示进行选择。出于多种原因，能够访问多种变体可能是有益的。例如，它允许您比较同一图像的不同版本，使您能够选择最适合您需求的版本。此外，它还可以帮助您避免创建重复的图像，这可能既耗时又无聊。

When it comes to image-generating AI, there is a seemingly limitless amount of variation available. One such example is the text-image generative AI Stable Diffusion, which provides users with four variations of an image at a time to choose from based on the text prompt. Having access to multiple variations can be beneficial for several reasons. For instance, it allows you to compare different versions of the same image, enabling you to select the best one for your needs. Additionally, it can help you avoid creating repetitive images, which can be time-consuming and boring.

查看多个变体可以帮助您探索创意过程的后续步骤。通过生成多个变体，您还可以通过将变体与其他构建块相结合来尝试不同的样式、效果和配色方案。这种方法对于艺术家、设计师和摄影师来说非常有用，他们可以在文本提示中附上自己创作的上传图像。生成式人工智能不仅可以向个人上传的照片、作品或原创艺术添加元素，而且许多还提供不断增长的滤镜库，从根本上改变它们的风格（图6-2）。许多生成式人工智能还提供放大功能，可以提高创意者上传的图像的质量和分辨率。

Reviewing multiple variations can support you in exploring the next steps your creative process will take. By generating several variations, you can also experiment with different styles, effects, and color schemes by combining variation with other building blocks. This approach can be useful for artists, designers, and photographers who can accompany a text prompt with an uploaded image of their own creation. Not only can generative AI add elements to a photo or work or original art that a person uploads, but many offer a growing library of filters to radically transform them stylistically (Figure 6-2). Many generative AI also offer the feature of upscaling that improves the quality and resolution of an image that a creative uploads.

A photograph and 2 animated versions of a garden. It is a response to generate fairies, heather, butterflies, and a pastoral scene using A I. The top one is a real photograph, the center one has butterflies, and the bottom one has fairies.

A photograph and 2 animated versions of a garden. It is a response to generate fairies, heather, butterflies, and a pastoral scene using A I. The top one is a real photograph, the center one has butterflies, and the bottom one has fairies.

上传您自己的照片并创建变体或添加新元素可以改变以前的工作，而使用图像编辑软件可能需要花费数小时的时间。在图6-2生成的图像中，请注意底部两张图像中水池的消失以及石南花和仙女的添加。

Uploading your own photos and creating variations or adding new elements can transform previous work in ways that might have taken many hours using image editing software. Within the images generated in Figure 6-2, notice the disappearance of the water pond in the bottom two images and the addition of heather and fairies.

用种子进行更微妙的变化实验

Experimenting with Seeds for More Subtle Variations

文本图像生成人工智能还创建称为“种子”的独特数字标识符，以允许对图像进行轻微排列。生成的每个图像都是一个种子。该种子允许您继续调整与该种子关联的图像。每个种子都与特定生成的图像有数字关联。并非所有生成式人工智能都会免费提供这些种子，但它们在增强原型方面非常有价值，只需稍加改动即可更好地符合您的愿景。当您生成一张您真正喜欢的图像，但其中可能有您不喜欢的部分而难以删除时，这非常有用。您可以重新生成相同的种子图像，并在提示中通过添加、减去或替换提示中的一个或多个单词来稍微修改文本（图6-3）。

Text-image generative AI also create unique numeric identifiers known as “seeds” to allow for slight permutations of an image. Every image generated is a seed. That seed allows you to continue to tweak the image associated with that seed. Each seed has a numeric association with a specific generated image. Not all generative AI provide these seeds freely, but they can be valuable in enhancing prototypes with slight variations to better align with your vision. This is useful when you generate an image you really like, but there may be one part of it you don’t like that you are having difficulty removing. You can regenerate the same seed image and, in your prompt, slightly modify the text by either adding, subtracting, or substituting one or more words of your prompt (Figure 6-3).

Two animated pictures for a response to generate a virtual world inside the head with cute robots, profiles, hyper-realistic, 3D, hyper-detailed, unreal game engine. The left one is a female cyborg. The right one is a female robot.

Two animated pictures for a response to generate a virtual world inside the head with cute robots, profiles, hyper-realistic, 3D, hyper-detailed, unreal game engine. The left one is a female cyborg. The right one is a female robot.

加减

Addition and Subtraction

当与人工智能交互时，加法和减法的构建块会被快速应用，并揭示任何生成内容的可供性和局限性。当您觉得缺少某些想要添加到文本、图像、音乐、视频或其他媒体的内容时，可以将添加构建块应用于提示。在提示中添加文本可能会促使人工智能生成不同的结果。然而，有时您会面临挑战，想要添加一些您的缪斯无法生成的仍然缺失的东西。根据您自己的技能和技术，您可能必须从缪斯手中接管控制权，并使用其他软件添加缺少的内容。这可能就像添加颜色一样简单，你根本无法让人工智能在 Photoshop 中生成或生成法学硕士拒绝创建的文本。

The building blocks of addition and subtraction are applied rapidly when interacting with an AI and reveal the affordances and limitations of any generative content. The building block of addition can be applied to a prompt when you feel something is missing that you’d like to add to text, image, music, video, or other media. Adding text to a prompt may provoke the AI to generate different results. However, sometimes you’ll be faced with the challenge of wanting to add something that is still missing that your muse cannot generate. Depending on your own skill and technique, you’ll likely have to take over control from your muse and add what is missing using other software. This may be as simple as adding colors you simply cannot get an AI to generate in Photoshop or generating the text that an LLM refuses to create.

当您觉得生成的内容中存在您想要删除的内容时，可以应用减法构建块。这可以是任何东西，例如生成的音乐片段开头的一层生成的声音对您不感兴趣，一个人的表现方式不符合您的审美或意图的图像，或者人工智能生成的文本你知道这是谎言或有偏见。您可以继续使用相同的提示生成内容，或者删除提示中可能导致 AI 添加您不需要的内容的单个单词。如果人工智能没有生成您想要的内容，那么您可能必须使用您熟悉的软件来启用您的技能和技术，例如 Photoshop 来裁剪图像或数字音频工作站来缩短人工智能生成的音轨，操纵某些使用均衡的频率，或将轨道用作多轨道合成中的一层。

The building block of subtraction can be applied when you feel something is present in generated content that you’d like to remove. This can be anything from a layer of generated sound in the beginning of a generated piece of music that doesn’t appeal to you, an image where how a person is represented doesn’t fit your aesthetic or intent, or text that an AI generates that you know is a lie or biased. You can continue to generate content with the same prompts or remove what might be a single word in your prompt that is causing the AI to add content you don’t want. If the AI is not generating what you want, then you may have to enable your skills and technique using software you are familiar with, such as Photoshop to crop an image or a digital audio workstation to shorten an AI-generated audio track, manipulate certain frequencies using equalization, or use the track as one layer in a multi-track composition.

我们可以使用加法工具来扩展文本-文本和文本-图像提示，或使用减法工具来缩短这些提示。这两种构建模块在创意产业中都很常见。每一个都可以应用于逐步开发想法和项目。使用法学硕士开发故事原型的作者可以通过逐渐构建他们的故事（一次一个场景）来使用加法。他们可能会从单个角色或场景开始，然后添加更多元素，如对话、冲突和解决方案，以创造丰富且引人入胜的叙事。相反，在收到反馈后，他们可能希望通过删除可能使其过于复杂的元素来重新生成故事的某些部分。

We can use the tool of addition to expand our text-text and text-image prompts or subtraction to shorten those prompts. Both building blocks are common across creative industries. Each can be applied to develop ideas and projects incrementally. Authors developing a prototype of a story using an LLM may use addition by gradually building up their story, one scene at a time. They might start with a single character or setting and then add more elements like dialogue, conflict, and resolution to create a rich and engaging narrative. Conversely, after receiving feedback they might want to regenerate some parts of the story by removing elements that might have made it overly complex.

视觉艺术中的添加可以指将新元素（例如颜色、形状、纹理或想法）融入到作品中。这个过程使艺术家能够尝试各种组件，将它们分层在一起以创造艺术品的深度和复杂性。视觉艺术家通常从简单的草图或概念开始，逐渐添加细节、层次和元素，以创造出更复杂、更具视觉吸引力的作品。例如，在绘画中，艺术家可能会想象一个基本的配色方案，当这些颜色在画布上显现并形成时，它们可能会激发新的冲动。艺术家直观或有意地改变、添加纹理并使用不同的颜色深浅来为不断发展的主题添加亮点。当使用文本图像生成人工智能时，提示的复杂性可能会增加。许多人工智能生成过于复杂的图像，需要删除特征。能够检查和审查生成的内容在很大程度上取决于艺术家为任何生成人工智能带来的技能和技术。在大多数情况下，艺术家的背景和经验将为虚拟画布本身的塑造提供信息，并产生更具吸引力和艺术性的结果，而生成式人工智能只能提供这些结果的开端。

Addition in visual art can refer to incorporating new elements, such as colors, shapes, textures, or ideas, into a composition. This process enables the artist to experiment with various components, layering them together to create depth and complexity in the artwork. Visual artists often begin with a simple sketch or concept and gradually add details, layers, and elements to create a more complex and visually appealing piece. In painting, for instance, an artist may imagine a basic color scheme, and as those colors manifest and take form on a canvas, they may inspire new impulses. Intuitively or intentionally artists vary, add textures, and use different shades of a color to add highlights to an evolving subject. When using a text-image generative AI, prompts can increase in their complexity. Many AI generate overly complex images that require features to be removed. Being able to inspect and review what is generated is highly dependent on the skill and technique that an artist brings to any generative AI. In most situations the background and experience of an artist will inform the shaping of the virtual canvas itself and result in far more compelling and artistic results that generative AI can only offer the beginnings of.

在音乐中，即兴钢琴家可以通过从单个音符开始，以不同的节奏重复该音符，通过组合这些音符开发出简单的旋律，分层低音线和和声来创作更丰富的音乐，从而自发地创作出一首作品。在不断发展的作品中添加和减去单个音符的简单性对于专注于声音生成的生成式人工智能来说要提示要困难得多。虽然它们可能擅长模仿特定艺术家的说唱的重复循环，但在生成的音乐中分离曲目或提示人工智能删除或减去您在整体声音中不想要的生成层要困难得多作品。这种艺术冲动需要定制的机器学习模型和大量的音频编辑，除非乐器轨道已经可分离，否则可能仍然无法减去您不希望在作品中出现的元素。参考我们原来的早餐提示，我们可以添加文本“菠菜和羊奶酪”来产生新的想法。

In music, improvising pianists can spontaneously generate a composition by just starting with a single note, repeating that note with varied rhythms, developing a simple melody by combining those notes, layering in a bass line and harmonies to create a richer piece of music. The simplicity of adding and subtracting individual notes in an evolving composition is much more difficult to prompt with generative AI that are focused on sound generation. While they may be good at emulating a repetitive loop for a rap that emulates a particular artist, it is much more difficult to separate tracks in generated music or prompt an AI to remove or subtract generated layers that you don’t want in the overall sound composition. That artistic impulse requires a customized machine learning model, and a lot of audio editing that may still not be able to subtract elements you don’t want present in the composition unless instrument tracks are already separable. Referring to our original breakfast prompt, we can add the text “with spinach and feta cheese” to generate new ideas.

作者：菠菜和羊奶酪的早餐有哪些创意？

Author: What are some breakfast ideas with spinach and feta cheese?

Muse：这里有一些选择：

Muse: Here are some options:

菜肉馅煎蛋饼：菜肉馅煎蛋饼是一种意大利煎蛋卷，可以填充多种配料。制作菠菜和羊乳酪菜肉馅煎蛋饼时，将菠菜放入平底锅中炒至变干；然后将其添加到打散的鸡蛋和碎羊奶酪的混合物中。在烤箱或炉灶上煮菜肉馅煎蛋饼直至凝固。

Frittata: A frittata is a type of Italian omelette that can be filled with a variety of ingredients. To make a spinach and feta frittata, sauté the spinach in a pan until wilted; then add it to a mixture of beaten eggs and crumbled feta cheese. Cook the frittata in the oven or on the stovetop until set.

鸡蛋烘焙：鸡蛋烘焙是将面包、鸡蛋、菠菜和羊奶酪分层放在烤盘中并在烤箱中烘烤而制成的菜肴。您可以使用任何类型的面包，也可以添加其他配料，例如西红柿丁、青椒或蘑菇。

Egg Bake: An egg bake is a dish made by layering bread, eggs, spinach, and feta cheese in a baking dish and baking it in the oven. You can use any type of bread, and you can also add other ingredients such as diced tomatoes, bell peppers, or mushrooms.

煎鸡蛋：煎鸡蛋是用菠菜和羊乳酪准备鸡蛋的另一种简单方法。只需将菠菜放入锅中炒至枯萎，然后在菠菜中心挖一个洞，然后将鸡蛋打入其中即可。上面放上碎羊奶酪，煮至蛋白凝固，但蛋黄仍然流淌。

Fried Eggs: Fried eggs are another simple way to prepare eggs with spinach and feta. Simply sauté the spinach in a pan until wilted, then make a hole in the center of the spinach, and crack an egg into it. Top with crumbled feta cheese and cook until the egg whites are set but the yolks are still runny.

炒鸡蛋：你也可以用菠菜和羊乳酪做炒鸡蛋。只需将菠菜放入锅中炒至枯萎，然后加入打散的鸡蛋，煮至凝固即可。上菜前在上面撒上碎羊奶酪。

Scrambled Eggs: You can also make scrambled eggs with spinach and feta. Simply sauté the spinach in a pan until wilted, then add in whisked eggs, and cook until set. Top with crumbled feta cheese before serving.

当生成图像的变体时，我们还会激活加法构建块，如图6-3中生成的图像。

We also activate the building block of addition when generating a variation of an image, such as the ones generated in Figure 6-3.

您也可以添加和减去您创建的图像的元素。图6-4显示了车道上一栋房屋的原始照片，其中有车辆和人员。提示应用了添加的构建块，将石南花添加到车道上，并将房屋放置在森林中。负面提示带走了图像中我们不想要的某些部分，在本例中是原始照片中的人和车辆。在某些生成人工智能中，您还可以添加键盘命令，例如使用括号来强调提示中您希望人工智能注意的单词，或者在文本图像提示中使用双括号以确保某些图像不会出现在生成的图像中。

A photograph and an animated version. It is a response to remove elements from the original photo and add elements instead of it using an anime-style filter. The top one is a real photograph and the bottom one is the animated one.

A photograph and an animated version. It is a response to remove elements from the original photo and add elements instead of it using an anime-style filter. The top one is a real photograph and the bottom one is the animated one.

代换

Substitution

创意人员经常应用替换。当将其应用于生成人工智能时，替代可能表现良好，也可能表现不佳。例如，您可能只想以特定艺术家的风格创建不同的对象，但无法获得一致的结果。在生成人工智能方面，替代是时好时坏的，因此坚持不懈并创造出多种变化是值得的。为此，如果替换的单词或短语给您带来不符合您的艺术视野的东西，您可能还必须牺牲您的一些愿景或意图。提示文本图像生成AI时也可以应用替换（图6-5）。

Creatives regularly apply substitution. Substitution may or may not behave well when it comes to applying it to a generative AI. For example, you may just want to create different objects in the style of a specific artist and not receive consistent results. Substitution is hit and miss when it comes to generative AI, so it pays to be persistent and create a number of variations. To that end you may also have to sacrifice some of your vision or intent if the substituted word or phrase gives you something that does not fit your artistic vision. Substitution can also be applied when prompting text-image generative AI (Figure 6-5).

Two animated pictures. It is a response to generate a photo-realistic image of a tree on the left and a photo-realistic image of a flower on the right.

Two animated pictures. It is a response to generate a photo-realistic image of a tree on the left and a photo-realistic image of a flower on the right.

请记住，每个生成的报价都是一个实验，可能会结合本章中强调的一些构建块。参与生成式人工智能就是与你的缪斯不断地来回对话；它正在理解可用的数据集，并试图找出（通过特定的匹配学习模型）如何充分利用您的提示来生成您可以使用的东西。

Remember that every generated offer is an experiment that will likely combine some of the building blocks highlighted in this chapter. Engaging with generative AI is a constant back-and-forth conversation with your muse; it is making sense of the data set made available to it and trying to figure out (through a specific matching learning model) how to make the most sense of your prompting to generate something you can use.

遮罩以替换图像的一部分

Masking to Substitute Parts of an Image

替代也可以是某些生成人工智能提供的功能。它们提供了进一步的灵感，即掩盖图像的某些区域的能力，并提示删除或替换图像的一部分以代替现有的图像（图6-6）。蒙版可让您识别要更改的图像的特定部分。一些生成式人工智能附带不同的画笔或其他工具来帮助您这样做。这可能是更改您生成的图像的一小部分及其所属种子的更有效方法。您还可以使用人工智能删除照片或您所做的其他创意作品的部分内容。

Substitution can also be a feature that is offered with some generative AI. They offer further inspiration in their capacity to mask certain areas of an image combined with a prompt to remove or substitute part of an image in place of what already exists there (Figure 6-6). Masks allow you to identify a certain part of an image that you want to change. Some generative AI come with different brushes or other tools to assist you in doing so. This may be a more effective way to change a small part of an image that you’ve generated and the seed in which it is a part. You can also use AI to take away parts of a photo or other creative work you have made.

Two animated pictures. It is a response to remove the cat from the portrait by using a mask in A I and substituting it with a dog. A masked cat face is present on the left and an edited photo with a dog face using Photoshop is present on the right.

Two animated pictures. It is a response to remove the cat from the portrait by using a mask in A I and substituting it with a dog. A masked cat face is present on the left and an edited photo with a dog face using Photoshop is present on the right.

迭代

Iteration

你可以利用变化、加法、减法和替换的构建模块走得更远。每次您完善提示以及提示对您可能提供的图像的影响程度时，您都在继续改进生成的内容，使其对您自己的创作过程有用。您创建文本、图像或其他媒体的迭代或版本，并且该迭代始终会有所不同，具体取决于生成式 AI 提供的功能，包括您希望原始生成内容更改的程度。用原型设计的语言来说，当你与人工智能打交道时，你可以以不同的形式创建不同版本或不同形式的想法表达。在许多情况下，您会逐步迭代内容，使其更接近您的设想。

You can go far with the building blocks of variation, addition and subtraction, and substitution. Every time you refine your prompts and the degree to which a prompt will affect an image you might offer, you are continuing to improve the generated content to make it useful to your own creative process. You create an iteration or a version of your text, image, or other media, and that iteration will always be different depending on the features a generative AI offers including the degree to which you want your original generated content to change. In the language of prototyping, you create different versions or representations of an idea in different forms when you engage with an AI. In many cases you move toward an iteration of your content that gets closer to what you envision.

粗略的原型可以生成为想法的要点列表，也可以是纸上的草图、猫的涂鸦、餐巾纸上的想法，或者只能用笔写在手上才能捕捉到的东西。当您开始开发最初的粗糙原型时，纸上的草图可以成为您制作的令人惊叹的绘画，然后在艺术画廊的开幕式上展示。该涂鸦变成了您与同事一起开发的视频游戏中的 3D 角色。餐巾纸上的想法变成了一个成熟的商业计划。你写在手上的那句话成为你创作的一首歌曲的副歌歌词，然后与你的乐队一起录制。

Rough prototypes can be generated as bullet point lists of ideas and can also be sketches on paper, a doodle of a cat, an idea on a napkin, or something you could only capture by writing it on your hand with a pen. When you start to develop your initial rougher prototypes, then the sketch on paper can become an amazing painting that you craft and then show at an art gallery opening. That doodle becomes a 3D character in a video game you are working on with colleagues. The idea on a napkin becomes a full-blown business plan. That sentence you wrote on your hand becomes the lyrics for a chorus of a song you came up with and then record with your band.

每一个简单、粗糙的原型，无论它是什么，都有可能提高保真度、分辨率和复杂性。生成式人工智能将支持您创建有针对性的内容，然后您可以对其进行完善。当您与缪斯互动时，您会意识到您的想法是由想法发展的早期阶段决定的。如果您将缪思返回给您的内容重新想象为您现在可以使用的粗略原型，则尤其如此。例如，您可以从法学硕士身上汲取想法，然后查看并编辑生成的文本内容。您的缪斯还可以生成图像、音乐、视频、代码、动画。尽管不完美或不完整、有问题、尴尬或不完全是您想要的，但生成的报价是构成正在进行的创意对话的一部分（图6-7）。

Every single, simple, rough prototype no matter what it is has the potential to increase in fidelity, resolution, and complexity. Generative AI will support you in creating targeted content that you can then refine. As you interact with your muse, you come to the realization that your ideas are informed by the earlier phases of idea development. This is especially true if you reimagine what your muse gives back to you as a rough prototype that you can now work from. You can bounce ideas off your LLM, for example, and review and then edit generated text content. Your muse can also generate images, music, videos, code, animation. Although imperfect or incomplete, buggy, awkward, or not exactly what you want, what is generated is an offer that forms part of an ongoing creative conversation (Figure 6-7).

Four animated pictures of an uncanny doll in a 2 by 2 pattern. Each version has some changes like different helmets, eyes, and costumes.

Four animated pictures of an uncanny doll in a 2 by 2 pattern. Each version has some changes like different helmets, eyes, and costumes.

有时，生成式人工智能给你带来的迭代会是不可思议的、尴尬的和奇怪的，但你也可能选择使用它。这些生成式人工智能的开发人员也可能会投入时间和资源来迭代算法和数据集，因此预计随着时间的推移，您将获得更有趣的结果和平台，为您提供更多功能，以创建更多版本的预期输出。图6-8已生成为3D 图像原型，以预先可视化它作为物理折纸雕像的外观。

An animated picture of a paper figurine. It is generated after the application of the Origami filter. It looks like a doll.

An animated picture of a paper figurine. It is generated after the application of the Origami filter. It looks like a doll.

增强

Augmentation

增强是一种用于放大某物或使某物更长或更宽的技术。想象一下您创建的画布，并且您想要预先想象如果可以通过内容、手势、场景和/或角色来补充您现有作品的长度或高度或两者扩展，它会是什么样子。生成式人工智能可以帮助您将其可视化。添加新元素怎么样？这也是有可能的。

Augmentation is a technique used to enlarge something or make something longer or wider. Imagine a canvas that you have created, and you wanted to pre-visualize what it might be like if it could be extended in length or height or both with content, gestures, scenes, and/or characters that complemented your existing work. Generative AI can support you in visualizing that. What about adding new elements? That too is possible.

借助一些生成式人工智能，这可以通过称为“outpainting”的过程来实现。这是深度学习中的一项生成任务，其中训练神经网络以生成超出作为输入数据提供的原始图像的新内容。它通常用于扩展或完成给定的输入图像、声音、视频或文本。这带来了令人兴奋的可能性。例如，您可以通过使用图像中视觉上已经存在的元素来扩展生成的图像，从而使整个画布更大。 OpenAI、Midjourney、Photoshop 和其他平台的 DALL-E 2 可以实现此功能。

With some generative AI, this can be achieved through a process called “outpainting.” It is a generative task in deep learning where a neural network is trained to generate new content beyond the original image that was provided as input data. It is often used to extend or complete a given input image, sound, video, or text. That allows for exciting possibilities. For example, you could extend an image that you generate by using elements already present visually in the image making the entire canvas bigger. This feature is possible with DALL-E 2 by OpenAI, Midjourney, Photoshop and other platforms.

Muse：DALL·E 2 是 GPT（生成式预训练变压器）-3 架构的 120 亿参数版本，已针对绘制任务进行了微调。给定描述所需图像或视频的提示，它会生成超出输入提示边界的相应新颖视觉内容。绘制任务要求模型理解输入提示的上下文，识别需要生成的内容，并创建连贯且真实的新内容。

Muse: DALL·E 2 is a 12-billion-parameter version of the GPT (Generative Pretrained Transformer)-3 architecture that has been fine-tuned for the outpainting task. Given a prompt describing a desired image or video, it generates a corresponding novel visual content that goes beyond the boundaries of the input prompt. The outpainting task requires the model to understand the context of the input prompt, identify what needs to be generated, and create new content that is both coherent and realistic.

外画可以被视为您自己视觉的创造性延伸，根据图像和文本输入生成新的视觉内容。模型生成的输出可用于各种应用，例如生成新的插图、绘画、动画等。以下面的数字集合为例。图6-9中的上图是人工智能想象作者在图书馆写一本关于人工智能的书的样子。中间和底部的图像是图书馆的扩展，其中包括也可能在图书馆中读写的其他人工智能。

A 3-part animated picture. The top picture represents an author writing a book on A I in a library. The middle and bottom pictures include one and two A Is, respectively, along with the author who may be reading and writing in a library. All are generated using A I.

A 3-part animated picture. The top picture represents an author writing a book on A I in a library. The middle and bottom pictures include one and two A Is, respectively, along with the author who may be reading and writing in a library. All are generated using A I.

缩小

Diminution

缩小是一种与增强相反的技术，旨在缩小手势的大小（例如，在音乐短语中），或者在生成内容的情况下，删除图像中您不想要的内容或更改比例或大小画布的。大多数图像生成人工智能更进一步，允许用户决定他们不想要图像中的内容。当该图像附有文本提示中的说明时，即可实现这一点。通过识别生成的图像中您不想要的内容，您可以引导您的灵感产生更理想的结果。例如，假设您想要一张高质量的人物照片，但不希望它看起来不切实际或扭曲。通过添加“不切实际”、“扭曲”或“额外的手指”等负面关键词，您可以引导生成式人工智能生成更逼真、更准确的图像。负面提示对于完善你的艺术风格也很有用。如果您对缪斯生成的结果不满意，请尝试添加一些否定关键字以引导其生成更符合您的艺术愿景的内容。此外，如果您想要裁剪图像的所有部分（除了一个部分），可以将缩小应用于生成的图像。之后，您可能会提示图像-图像人工智能进一步迭代该部分或升级它。我们不可避免地会期望生成人工智能图像平台添加新功能，允许用户放大图像并附带文本提示。

Diminution is a technique opposite to augmentation and meant to reduce the size of a gesture (e.g., in a musical phrase) or, in the case of generated content, remove something in the image you don’t want or change the scale or the size of the canvas. Most image generative AI take it a step further by allowing users to decide what they don’t want in an image. This is achieved when that image is accompanied by instructions in a text prompt. By identifying what you don’t want in your generated image, you can guide your muse toward producing more desirable results. For example, let’s say you want a high-quality photograph of a person, but you don’t want it to look unrealistic or distorted. By adding negative keywords like “unrealistic,” “distorted,” or “extra fingers,” you can steer the generative AI toward producing a more lifelike and accurate image. Negative prompts can also be useful in refining your artistic style. If you’re not satisfied with the results your muse generates, try adding some negative keywords to guide it toward producing something more aligned with your artistic vision. Additionally, diminution can be applied to a generated image should you want to crop all parts of that image except one. After, you might prompt an image-image AI to further iterate on that one part or upscale it. Inevitably we can expect new features to be added to generative AI image platforms that will allow users to zoom in to an image with an accompanying text prompt.

换位

Transposition

最后一点可以在下面的示例中详细说明。

This last point can be detailed in the example that follows.

生成的一系列戴着兜帽、披着斗篷、手持光剑的猫经历了数百次迭代，将提示从一个生成人工智能添加、减去和转置到另一个生成人工智能，以获得如图6-10所示的所选变体。有时猫没有胡须。其他时候，兜帽遮住了整个头部。进行多次迭代具有战略意义，因为它还可以让您完善提示和负面提示。在图6-10的情况下，提示至少在提示开始时经历了以下迭代。每个提示描述后面都附有风格参考：

• 穿着斗篷、拿着光剑的猫

• Cat wearing a cloak and holding a light sabre

• 头上披斗篷的猫拿着光剑

• Cat with a cloak over its head holding a light sabre

• 穿着斗篷的猫拿着光剑

• Cat in a cloak holding a light sabre

• 穿着连帽衫、拿着光剑的猫

• Cat in a hoody with a light sabre

• 拿着光剑的僧帽猫

• Cat in a monk’s hood holding light sabre

• 披着连帽斗篷、拿着光剑的猫

• Cat with a light sabre in a hooded cloak

• 戴着光剑的猫

• Cat in a hood with a light sabre

An animated picture of a cat. It is a response to generate a cat in a hood with a lightsaber using A I.

An animated picture of a cat. It is a response to generate a cat in a hood with a lightsaber using A I.

尽管需要添加“肖像镜头”和“全身”等词语以确保至少显示猫身体的一部分，而不仅仅是头部，但最终的提示产生了最佳结果。

The final prompt produced the best results even though words like “portrait shot” and “full body” needed to be added to ensure at least part of the cat’s body was shown, not just the head.

由于结果并不完全是我想要的，我决定在下一组提示中强调光剑，并在具有特定风格过滤器的图像-图像 AI 中使用图6-10 。结果如图6-11所示，除了我选择了黑白风格的滤镜之外，它在所有方面都非常接近。到目前为止，这是我能够达到预期提示的最接近的结果，但我决定现在在同一个图像-图像生成人工智能中使用该图像，看看我会得到什么。

A greyscale animated picture of a cat. It is a response to generate a cat in a hood, holding a lightsaber.

A greyscale animated picture of a cat. It is a response to generate a cat in a hood, holding a lightsaber.

图6-11 生成结果尽管图6-12中的猫是用两只爪子生成的，但我还是选择了这个图像。此外，猫的渲染还需要额外的步骤，这意味着 GAN 在渲染最终的猫之前能够来回 17 次。

A hyper-realistic, greyscale animated picture of a cat. It is a response to generate a cat in a hood, holding a lightsaber.

A hyper-realistic, greyscale animated picture of a cat. It is a response to generate a cat in a hood, holding a lightsaber.

图6-12中的猫随后被用于图像-图像 AI 中，激发了一种新想法，即生成穿着连帽斗篷、携带点燃的灯笼而不是光剑的猫。在图像附带的提示中，“光剑”一词被“灯笼”一词取代。完整的提示内容如下：“穿斗篷的猫携带灯笼超现实虚幻引擎 3d 黑白”。添加了一些风格过滤器，但很难获得所需的外观和感觉。如图6-13所示，没有灯笼，猫眼有些诡异；这只猫正处于某种令人毛骨悚然的墓地里，头巾现在已经脱落了。此时，您将有多种选择，包括使用图6-13（如果您碰巧喜欢它）。您还可以像我一样使用否定提示，以确保在人工智能下次生成尝试后不会出现图像的特定部分。

A greyscale animated picture of a cat. It is a response to generate a cat in a cloak, carrying a lantern hyper-realistic unreal engine.

A greyscale animated picture of a cat. It is a response to generate a cat in a cloak, carrying a lantern hyper-realistic unreal engine.

图6-13的持续迭代包括添加“鹅卵石街道”一词，并在特定的生成式 AI 平台中使用括号强调某些词。图6-14是经过二十多次迭代后的结果。

A 3 D animated picture of a cat. It is a response to generate a cat in a cloak carrying a lantern cobblestone path hyper-realistic unreal engine.

A 3 D animated picture of a cat. It is a response to generate a cat in a cloak carrying a lantern cobblestone path hyper-realistic unreal engine.

沉淀在图6-14中斗篷猫的图像上，该图像集合被搁置并接受了新的图像。图6-15是电气实验室中猴子进行数十次迭代的结果，将文本提示和图像从一种生成式 AI 转换为另一种。

A 3 D animated picture of a monkey. It is a response to generate a hyper-realistic monkey in an electric lab, an unreal game engine, cute, and cuddly.

A 3 D animated picture of a monkey. It is a response to generate a hyper-realistic monkey in an electric lab, an unreal game engine, cute, and cuddly.

在这个创作关头，换位法意外地发挥了作用。之前的提示“披着斗篷的猫带着灯笼鹅卵石小路超现实虚幻引擎3D黑白黄”伴随着图6-15，电气实验室中的猴子，产生了新的形象。当你尝试生成式人工智能时，值得注意的是，你最终可能会得到比你想象的更好的东西，尤其是当你接受错误或意外内容时。

At this creative juncture, transposition came into play by accident. The previous prompt of “cat-in-a-cloak-carrying-a-lantern-cobblestone-path-hyper-realistic-unreal-engine-3D-black-and-white-with-yellows” accompanied Figure 6-15, the monkey in an electric lab, resulting in a new image. What is remarkable when you experiment with generative AI is that you may end up with something better than what you imagined, especially when you embrace mistakes or unintended content.

A hyper-realistic 3 D animated picture of a cat. It is a response to generate a cat in a cloak carrying a lantern cobblestone path unreal engine.

A hyper-realistic 3 D animated picture of a cat. It is a response to generate a cat in a cloak carrying a lantern cobblestone path unreal engine.

提示与响应

Prompt and Response

生成式人工智能的另一个常见特征是音乐中常见的构建模块。

Another building block that is a familiar feature of generative AI is commonly practiced in music.

在不同的音乐传统中，“呼唤和回应”的实践是指一系列不同的音乐或节奏手势，其中第二个手势被视为对第一个手势的直接评论或响应。这种情况经常发生在非洲音乐、爵士乐和布鲁斯音乐中，但也可以在许多其他音乐流派中找到。 “呼叫”可以是一位音乐家演奏的任何定义的或即兴的手势，而“响应”可以是另一位音乐家或一组音乐家的直接重复、变化或新手势形式的答案。

In different musical traditions, the practice of “call and response” refers to a succession of two distinct musical or rhythmic gestures where the second gesture is heard as a direct commentary on or response to the first. This happens often in African music, jazz, and blues but can be found across many other musical genres. The “call” can be any defined or improvised gesture played by one musician, and the “response” can be a direct repetition, a variation, or an answer in the form of a new gesture from another musician or group of musicians.

在生成人工智能的背景下，“呼叫”对应于用户的输入或提示。这可以是向 ChatGPT-4 等语言模型发出的特定提示，也可以是作为起点提供给人工智能的图像或视频。

In the context of generative AI, the “call” corresponds to the user’s input or prompt. This could be a specific prompt given to a language model like ChatGPT-4 or an image or video provided to an AI as a starting point.

“响应”对应于生成式人工智能产生的输出。此输出直接受到用户输入的影响，但这一次，报价是由人工智能创建的。例如，法学硕士根据给定的提示生成文本，或者音乐生成人工智能可能会根据输入片段的灵感创作一首新的音乐。

The “response” corresponds to the output produced by the generative AI. This output is directly influenced by the user’s input, but this time, the offer is created by the AI. For example, an LLM generates text that is based on the given prompt, or a music-generating AI might create a new piece of music inspired by the input piece.

这种交互允许动态和迭代的过程，就像音乐一样。用户可以根据人工智能的“响应”调整他们的“呼叫”，从而实现可以带来意想不到的创造性结果的来回。通过这种方式，用户和人工智能进行协作，共同为最终的原型做出贡献。

This interaction allows for a dynamic and iterative process, much like in music. The user can adjust their “call” based on the AI’s “response,” enabling a back-and-forth that can lead to unexpected and creative results. In this way, the user and the AI collaborate, each contributing to the final prototype.

对于法学硕士，您可以用文本提示它，然后它会生成内容作为对您提示的响应。如果您不喜欢第一个内容，您可以重新生成更多内容，或者根据生成的内容细化提示。人工智能生成的内容是出乎意料的，它会告诉你如何回应人工智能。这种循环的对话会产生一个你可能不必开始的想法。图6-17中的图像是代表相同提示“手握大脑的机器人”的拼贴画，在不同的文本图像生成人工智能中进行转换，并进行细化以更接近底部手握大脑的机器人的实际图像正确的。除了文字提示之外，在第一个生成的图像（其本身是由作者伸出的手和手臂的照片提示）之后，所有提示都包含了生成的图像。当我们收到越来越有用的原型时，我们与人工智能的互动越多，我们提示它们的方式就越改进。

Six animated pictures of a robot in a 3 cross 2 pattern. It is a response to generate a robot with the brain in hand. Each picture has some variations in the placement of the brain. The bottom 2 pictures have a hand and brain only.

Six animated pictures of a robot in a 3 cross 2 pattern. It is a response to generate a robot with the brain in hand. Each picture has some variations in the placement of the brain. The bottom 2 pictures have a hand and brain only.

当您开始与生成人工智能缪思互动时，您将开始开发自己的构建块、策略和方法来生成原型，这将有助于您自己的创作过程。这样做会让你内心的创造者向许多不同的方向旋转，而这可能正是你摆脱创意障碍所需要的（图6-18）。可以生成许多原型，并且在与缪斯互动的过程中您会发现无尽的想法。事实上，仅仅与一个人接触就会加速你自己的创作过程，并扩展你的想象力，让你考虑以前可能从未想到过的新想法。

As you begin to interact with your generative AI muse, you will begin to develop your own building blocks, strategies, and methods to generate prototypes that will be helpful to your own creative process. Doing so will spin your inner creator in many different directions, and that might be just what you need to get out of a creative block (Figure 6-18). Many prototypes are possible to generate, and you will find endless ideas in the act of engaging with your muse. In fact, just engaging with one will accelerate your own creative process and expand your imagination for you to consider new ideas that you might never had thought of before.

An animated picture. It is a response to generate a comic strip with targeted A I filters and neural filters in Photoshop. It also includes some text descriptions on A I. The last text reads, it's like having a really smart and helpful friend.

An animated picture. It is a response to generate a comic strip with targeted A I filters and neural filters in Photoshop. It also includes some text descriptions on A I. The last text reads, it's like having a really smart and helpful friend.

致谢

Acknowledgments